Quantitative Methods in Cognitive Semantics:

Corpus-Driven Approaches
Cognitive Linguistics Research
46

Editors
Dirk Geeraerts
John R. Taylor
Honorary editors
René Dirven
Ronald W. Langacker

De Gruyter Mouton
Quantitative Methods in
Cognitive Semantics:
Corpus-Driven Approaches

Edited by
Dylan Glynn
Kerstin Fischer

De Gruyter Mouton
ISBN 978-3-11-022641-6
e-ISBN 978-3-11-022642-3
ISSN 1861-4132

Library of Congress Cataloging-in-Publication Data

Quantitative methods in cognitive semantics: corpus-driven approa-

ches / edited by Dylan Glynn, Kerstin Fischer.
p. cm.
Includes bibliographical references and index.
ISBN 978-3-11-022641-6 (hardcover : alk. paper)
1. Semantics. 2. Cognitive grammar. 3. Computational linguis-
tics. I. Glynn, Dylan. II. Fischer Kerstin, 1966⫺
P325.Q36 2010
4011.430721⫺dc22
2010038790

Bibliographic information published by the Deutsche Nationalbibliothek

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie;
detailed bibliographic data are available in the Internet at http://dnb.d-nb.de.

쑔 2010 Walter de Gruyter GmbH & Co. KG, Berlin/New York

Printing: Hubert & Co. GmbH & Co. KG, Göttingen
⬁ Printed on acid-free paper
Printed in Germany
www.degruyter.com
Table of contents

Corpus-driven Cognitive Semantics.

Introduction to the field ……………………………………………... 1
Dylan Glynn

Quantitative methods in Cognitive Semantics.

Introduction to the volume ………………………………………….. 43
Kerstin Fischer

Section I Corpus methods in Cognitive Semantics

The doctor and the semantician ……………………………………... 63
Dirk Geeraerts

Balancing Acts: Empirical pursuits in Cognitive Linguistics ….……. 79

John Newman

Does frequency in text instantiate entrenchment

in the cognitive system? ………………………………………….….. 101
Hans-Jörg Schmid

Section II Advancing the science: Theoretical questions

The aspectual coercion of the English Durative Adverbial …………. 137
Stefan Fuhs

The force dynamics of English complement clauses:

A Collostructional Analysis ……………………………………….… 155
Martin Hilpert

Accounting for the role of situation in language use

in a Cognitive Semantic representation of sentence mood ………….. 179
Kerstin Fischer

Exemplars and analogy:

Semantic extension in constructional networks ………………….….. 201
Arne Zeschel
v Table of contents
vi

Section III Advancing the science: Methodological questions

Marrying cognitive-linguistic theory and corpus-based methods:

On the compositionality of English V NP-idioms …………………... 223
Stefanie Wulff

Testing the hypothesis.

Objectivity and verification in usage-based Cognitive Semantics ….. 239
Dylan Glynn

Beyond the dative alternation:

The semantics of the Dutch aan-Dative ………………………..……. 271
Timothy Colleman

Corpus-based evidence for an idiosyncratic

aspect-modality relation in Russian …………………………………. 305
Dagmar Divjak

Section IV Towards an empirical Cognitive Semantics

Quantitative approaches in usage-based Cognitive Semantics:

Myths, erroneous assumptions, and a proposal ……………………… 333
Stefan Th. Gries and Dagmar Divjak

Empirical cognitive semantics: Some thoughts ………………...….. 355

Anatol Stefanowitsch

Author index ………………………………………………………… 381

Subject index ………………………………………………………… 389

List of Contributors

Timothy Colleman Martin Hilpert

Dutch Linguistics Department School of Language and Literature
Ghent University Freiburg Institute for Advanced
[email protected] Studies (FRIAS)
[email protected]
Dagmar Divjak
Russian and Slavonic Studies John Newman
University of Sheffield Department of Linguistics
[email protected] University of Alberta
[email protected]
Kerstin Fischer
Institute for Business Communication Hans-Jörg Schmid
and Information Science Institut für Englische Philologie
University of Southern Denmark Ludwig-Maximilians-Universität
[email protected] München
[email protected]
Stefan Fuhs muenchen.de
Institut fürAnglistik/Amerikanistik
University of Jena Anatol Stefanowitsch
[email protected] Department of Language Literature
and Media, University of Hamburg
Dirk Geeraerts anatol.stefanowitsch@uni-
Department of Linguistics bremen.de
University of Leuven
[email protected] Stefanie Wulff
Department of Linguistics and
Dylan Glynn Technical Communication
Centre for Language and Literature University of North Texas
University of Lund [email protected]
[email protected]
Arne Zeschel
Stefan Th. Gries Institute for Business Communication
Department of Linguistics and Information Science
University of California University of Southern Denmark
Santa Barbara [email protected]
[email protected]
Corpus-driven Cognitive Semantics
Introduction to the field

Dylan Glynn

1
Is quantitative empirical research possible for the study of semantics? More
specifically, can we use corpus data to produce testable and falsifiable re-
sults for the description of meaning? If so, what are the advantages and
what are the limitations of such an approach? Such questions are often
heard and it would seem they are behind much of the reticence and scepti-
cism for corpus-driven techniques in the broader Cognitive Linguistic
community. This volume grew out of the workshop Usage-Based and
Quantitative Methodology in Cognitive Semantics at the International Cog-
nitive Linguistics Conference 2007, in Kraków, where the participants were
asked to answer these very questions. This book both critiques and supports
the application of such techniques to the study of meaning in language.

1. Cognitive Linguistics – A necessarily empirical approach

Cognitive Linguistics is beginning to realise the implications of its own

theoretical framework. When Langacker (1987, 1988) first outlined the
theory of entrenchment and his usage-based model, he may not have ap-
preciated the extent to which these proposals would be implied in the trends
of 21st century linguistics. Indeed, Langacker’s theoretical proposal argu-
ably set the stage for a major paradigm shift in linguistics, from theory-
driven to empirical research. Similarly, when Lakoff (1982, 1987) argued
that we need to approach language holistically, that categorisation based on
human experience is the foundational structure of meaning, he is likely to
have underestimated the implications of this move. A holistic approach to
meaning goes far beyond adding pragmatics to semantic analysis – it entails
that everything speakers know of the world, including the whole domain of
sociolinguistics, is necessarily drawn into the mainstream of language de-
scription.
It is due to these theoretical tenets that the roots of Cognitive Linguistics
were, and inevitably so, an empirical movement seeking to account for the
2 Dylan Glynn

totality of language as a socio-cognitive phenomenon. The non-empirical

emphasis of earlier research can be attributed to methodological practices
inherited from the Structuralist and Generative (or Mentalist) schools.
These schools assume an object of study based on society’s langue or the
ideal speaker’s competence, and as such, their methods can rely exclusively
on introspection.
Cognitive Linguistics’ model of language necessarily entails empirical
methodology. In fact, Fillmore (1985), Talmy (1985), Lakoff (1987), and
Langacker (1987) can be said to have produced one of the first data-
oriented models of language. Their model dismantles a hundred years of
fervent linguistic theorising. They deconstruct the Structuralist distinction
of semantics and pragmatics, demonstrate the arbitrariness of the Modular-
ist division between lexis and syntax, and argue that language structure is a
result of usage, based on general cognitive capacities. The object of study is
finally language use in all its complexity, rather than langue or ‘compe-
tence’, lexis or syntax, and semantics or pragmatics.
The model of language proposed by Cognitive Linguistics is so com-
pletely simple that it places the emphasis squarely on method and data.
Rather than simplifying the object of study by carving off its complexities
with hypothetical modules of language structure, it lands the linguist in the
midst of a chaotic phenomenon that is the nature of all socially structured
systems. The original cognitive theoreticians did this in dialectic with
Structuralism and Mentalism. Most linguists today would agree that the
Cognitivists succeeded in demonstrating how such reductionist theories
failed. But did the Cognitivists offer a viable alternative? In some sense,
they took us out of the frying pan and into the fire. Firstly, claiming that
grammar is semantics-driven and that linguistic meaning extends beyond
traditional semantics to all world knowledge exponentially increases the
complexity of the object of study. Secondly, both Structuralism and Men-
talism claimed that early Cognitive Linguistics offered no means for testing
hypotheses or falsifying analyses. They were right. Lakoff’s (1987) claim
that co-occurrence of semantic phenomena is the basis of Cognitive Lin-
guistics’ scientific method is largely vacuous unless results obtained with
this method can be falsified. Likewise, it is all very well for Langacker
(1987) to argue that grammaticality is relative and varies from person to
person. But if this is the case, how can we test proposed grammatical struc-
tures?
Despite the fact that neither Lakoff nor Langacker developed an empiri-
cal line of research, together their proposals establish the cornerstone of
 Corpus-driven Cognitive Semantics 3

true empirical linguistic enquiry. Lakoff argued that evidence for his propo-
sals comes from the co-occurrence of linguistic phenomena. Reframed, his
argument was for inductive analysis, a method of analysis which is the
norm in the social sciences. Making generalisations based on a sample, then
extrapolating those generalisations to the population is the basis of induc-
tive scientific research, and the only viable method for the social sciences.
Yet, without rigorous and testable techniques for establishing how repre-
sentative or reliable a generalisation is, the ‘results’ remain merely hypoth-
eses based on very small samples. Indeed, as was typical at the time, a
sample was no more than the internal language knowledge of the linguist
and, perhaps, a few colleagues.
It is here that Langacker’s theory of entrenchment and the usage-based
model step in – they offer an operationalisation of grammar and grammati-
cality. In empirical research, operationalisation is a basic analytical tool –
the definition of an object of study by how one can measure it. Once deal-
ing with a measurable object of study, one may test the accuracy of its de-
scription and falsify predictions made about it. The ability to falsify results
is crucial if language analysis is to progress without grammaticality tests.
Langacker’s theory allows us, therefore, to test the accuracy of generalisa-
tions based on Lakoff’s identification of co-occurring linguistic phenom-
ena. These two points are crucial for quantitative corpus-driven semantic
research and we return to these points below.
Having established that the theory of Cognitive Linguistics is inherently
empirical, can we say, however, that the paradigm is empirical in practice?
To some extent, empiricism, be that based upon the experimental data of
psycholinguistics or upon the collected data of corpus linguistics, is a new
turn for Cognitive Linguistics. Yet, despite the recent expansion of empiri-
cal methods, both corpus-driven and experimental research go back to the
2
origins of Cognitive Linguistics.
Since the very beginning of the movement, a small yet important num-
ber of scientists have taken very seriously the methodological implications
of the cognitive linguistic model of language. Two early works, by scholars
whose names were crucial in the foundational years of the paradigm, are
testimony to this. Firstly, Dirven et al. (1982) is a corpus-driven study of
depth and maturity that would hold its own today. Secondly, Lehrer (1982)
is a linguistic experiment whose design and thoroughness exceed many of
our current studies. The former investigation led, directly and indirectly, to
the development of corpus Cognitive Linguistics in Belgium and Germany.
The latter work, recently republished by Oxford University Press (Lehrer
4 Dylan Glynn

2009), is enjoying renewed attention a quarter of a century later. Although

the experimental track is equally important to the development of Cognitive
3
Linguistics, it lies beyond the purview of this volume. The current volume
belongs to the tradition, established with Dirven et al. (1982), focusing on
methods employing corpus data.
From Dirven et al. (1982), via Dirven & Taylor (1988), Schulze (1988),
Rudzka-Ostyn (1989, 1995, 1996) and Goossens (1990) up until Geeraerts
(1993b), Schmid (1993, 1997, 2000) and Geeraerts et al. (1994, 1999),
corpus-driven Cognitive Linguistics grew slowly but surely. During the
1990s, the movement was given added weight by the demonstration of the
inadequacies of the analytical techniques current at the time. The Structur-
alist and Mentalist-Modularist assumptions about how language works,
assumptions that still drove the analytical techniques of early Cognitive
Linguists, were shown to be ineffective. Geeraerts (1993a) demonstrated
theoretically that the study of polysemy still made Structuralist assumptions
about how meaning is organised. Essentially, the kind of radial network
analysis popularised by Lakoff (1987) sought to identify discrete lexical
senses where there was no reason to suppose that such discrete senses exist:
The tremendous flexibility that we observe in lexical semantics suggests a
procedural (or perhaps ‘processual’) rather than a reified conception of
meaning; instead of meanings as things, meaning as a process of sense cre-
ation would seem to become our primary focus of attention. Geeraerts
(1993a: 260)
Two years later, Sandra & Rice’s (1995) experimental study ‘Mirroring
whose mind, the language user’s or the linguist’s demonstrated that the
techniques used to study polysemy in Cognitive Semantics were, in Pop-
per’s sense, ad hoc. Just as the approaches in Cognitive Linguistics had
inherited the analytical construct of Structuralism - senses come in discrete
categories - they had also inherited that the analytical technique of the
Mentalist grammars – introspection can identify grammaticality. However,
for Cognitive Linguistics, with no ‘ideal speaker’, the intuition of one na-
tive speaker (the linguist) cannot be assumed to represent the language.
Introspection is still widely used in Cognitive Linguistics, but following the
publication of Sandra & Rice’s study, descriptive semantic analysis based
entirely on introspection, lost much scientific credibility within the cogni-
tive community.
This critique of introspection as the sole basis of linguistic investigation,
in combination with the establishment of Cognitive Linguistics as an inde-
pendent model of language, led to the growth in experimental and corpus-
 Corpus-driven Cognitive Semantics 5

driven methods. To understand, however, the extent to which Cognitive

Linguistics has always been a empirical approach to language, let us briefly
revisit the foundational theories of Fillmore (1985), Talmy (1985), Lan-
gacker (1987), and Lakoff (1987).

2. Usage-based Cognitive Linguistics - unifying Mentalist and

Structuralist theories, operationalising meaning and grammar

Geeraerts (2006a: 398-415, 2006b, 2007, 2010) argues that Cognitive Lin-
guistics has recontextualised language description, returning elements that
were removed by Structuralism and Mentalism-Modularism. More specifi-
cally, he shows that the tenets of the paradigm entail that meaning drives
grammar. This meaning, he stresses, is conceptual, emergent, and social.
Geeraerts’ idea of the socio-conceptual recontextualisation of Cognitive
Linguistics is difficult to dispute, seeing the current trends. The basic ar-
gument is that there has been a linear progression from Structuralist and
then Mentalist-Modularist decontextualisation to the recontextualisation of
grammar in Cognitive Linguistics. However, one could perhaps go further
and argue that this is but a symptom of a more fundamental unification of
language theories. There are two issues here - how the theory of Cognitive
Linguistics unites Mentalist and Structuralist language models, and how
this unification operationalises grammar and semantics. As such, it can be
said that a combination of the usage-based model and corpus-driven re-
search resolves some of the great debates of 20th century linguistics.

2.1. Entrenchment – operationalised grammar

Langacker’s (1987) theory of entrenchment unites the two dominant 20th

century models of language. If we read Langacker’s position in Chomskyan
4
terms, the idea of an individual’s ‘competence’ is maintained. Each
speaker possesses a mental grammar with intuitions about correctness in
language. However, for Langacker, this individual mental grammar does
not represent a language. In other words, there is no ideal speaker - each
individual’s grammar is a learnt code that is unique to the extent that the
individual’s exposure to language is different from other individuals’ in the
speech community. It is this point that is crucial since it brings us to de
Saussure’s langue. The speaker, in his or her use of language, constantly
6 Dylan Glynn

judges a perceived set of usage norms, which in Structuralist times, would

have been called langue.
Entrenchment, however, is more than a theory of grammaticality, it is an
operationalisation of grammaticality. Langacker is primarily concerned
with the status of a linguistic unit, but in the usage-based model, entrench-
ment can be extended to replace the notion of grammatical correctness,
where the principle of frequency of use for the individual is extended to
that of frequency of occurrence in the community. This operationalisation
defines the phenomenon of grammar by showing how one may observe and
measure it.
Every use of a structure has a positive impact on its degree of entrenchment,
where […] disuse has a negative impact. […] Moreover, units are variably
entrenched depending on the frequency of their occurrence (driven, for ex-
ample, is more entrenched than thriven)… The absence of a sharp division
between units and nonunits has the consequence that the scope of a gram-
mar is not precisely delimited. (Langacker 1987: 59-60).
Here, Langacker offers an observable and quantifiable definition of an in-
dividual’s grammar and demonstrates how it relates to the grammar of a
5
language.
The measurable definition of grammar proposed by Langacker unifies
the internal Mentalist and the external Structuralist conceptions of lan-
guage. This unification eliminates the need for both the langue-parole and
competence-performance distinctions. It restores the holistic complexity of
language analysis by treating the emergent structure of grammar as a result
of its use by individuals in a community. Seen in this light, not only is
Cognitive Linguistics recontextualising language as a social phenomenon,
it is uniting the two principal 20th century models of language and defining
its framework in such a way as to permit empirical research.

2.2. Categorisation - operationalised encyclopaedic semantics

Entrenchment, as a theory, bridges Mentalism and Structuralism. It is,

however, a symbolic theory of grammar, based on the pairing of form and
meaning. Thus, semantics is crucial and basic to any analysis that employs
it. Can we also operationalise meaning? Stefanowitsch (this vol.) asks this
very question and offers four operational definitions. However, before we
ask that question, we must emphasise that meaning is not restricted to
propositional or referential semantics in Cognitive Linguistics. Lakoff’s
 Corpus-driven Cognitive Semantics 7

(1987) theoretical work represents the foundations of Cognitive Semantics

and his position mirrors Langacker’s unification of langue and ‘compe-
tence’ in that it also dismantles hypothetical sub-parts of language struc-
ture, such as linguistic semantics and pragmatics. Following Fillmore’s
(1985) work, Lakoff argues for a non-propositional semantics based on
world knowledge – encyclopaedic semantics.
Yet, encyclopaedic semantics should not be seen as merely a combina-
tion of linguistic semantics and pragmatics; it places the holistic complexity
of human experience and semiosis in the driving seat of language produc-
tion. The study of language must, therefore, account for all that we know
about the world from denotational reference, through cultural knowledge to
even social expectations. In concrete terms, any study must simultaneously
account for the semantic motivation behind and interaction between syntax,
morphology, lexis, prosody, and all of this relative to discourse structures,
world knowledge, and social variation. Moreover, this sea of infinite com-
plexity varies subtly from individual to individual. It is the sum of all this
that represents our object of study – meaning. The seeming impossibility of
scientifically describing such multidimensional complexity is what led
Structuralism to treat language use separately, the parole, and what led
Mentalism-Modularism to dismiss it entirely. However, for a non-modular
semantically driven theory of language, the complexity of language use is
the basis of grammar. Can such a broad-reaching understanding of meaning
be operationalised scientifically? Lakoff attempts this through the notion of
conceptual categorisation.
Categorisation is a symbolic distinction between difference and simi-
larity. This notion permits Lakoff (1987) to operationalise meaning. Things
that are conceived as similar are grouped together, distinct from things that
are not similar. The sets of things conceived of as similar are called con-
cepts. This use of the notion concept is a powerful operationalisation of
meaning and allows us, as linguists, to use the same analytical tool to ac-
count for the full complexity of encyclopaedic semantics. It explains why
two physically different chairs are both labelled chair, why the lexeme bill
is not ambiguous in a restaurant scenario, why she sneezed the napkin off
the table is comprehensible, but also how conceptual metonymy and meta-
phor function.
The idea that, as cognitive beings, we are constantly judging similarity
and dissimilarity, that through this ability we order the chaotic and dynamic
environment in which we live, and that this ability and process is basic to
language seems reasonable. Moreover, at a theoretical level, it is a reason-
8 Dylan Glynn

able operationalisation of meaning. But does it work? Can this abstract

definition lead to quantification and measurement? In fact, it is already the
basis of much corpus-driven research. Co-occurrence and correrlation are
fundamental to corpus investigation. Whether it is formal co-occurrence,
such as that indentified by collocation in Collostructional Analysis, or se-
mantic co-occurrence, such as that indentified in the usage-feature analysis
of the behavioural profile approach, corpus research functions by identify-
ing sets of similar things. This is exactly Lakoff’s understanding of concep-
tual categorisation. So we see that just as frequency can operationalise
grammaticality, co-occurrence can operationalise categorisation.
Thus, we can say that frequency of co-occurrence, which is fundamental
to corpus research, is a quantitative operationalisation of the basic theories
of Cognitive Linguistics – entrenchment and categorisation. These theories,
entrenchment and categorisation, explain grammar and meaning. However,
co-occurrence must be understood as much more than formal co-
occurrences. As stressed above, we must simultaneously account for the
interaction of all dimensions of meaning in order to explain language. But it
is precisely this daunting task that quantitative corpus-linguistics is best
placed to achieve. Let us consider how.

2.3. Linguistic complexity - A multifactorial approach to language

We have so far established that Cognitive Linguistics has freed linguistic

research of complex theoretical models, the models that were designed to
tie down and render the complexity of natural language simple enough to
study rigorously. We have also seen how grammar and meaning are ac-
counted for theoretically and how these accounts are operationalised in
terms that are applicable to corpus-driven research. Yet still, how is it pos-
sible to scientifically account for such an immense and complex linguistic
system, a system that varies from speaker to speaker and from context to
context? With no core-grammar, no langue, not even propositional or refer-
ential semantics, what hope do we have of capturing, accurately and rigor-
ously, the conceptual structure that we believe motivates language? Is it
truly possible to make generalisations and write grammars without reducing
this complexity? One answer lies in multifactorial modelling. Of all the
advantages of corpus-driven research, this stands out as one of the most
important. Geeraerts (2006a) summarises succinctly the basic advantage:
 Corpus-driven Cognitive Semantics 9

Corpus materials provide the firm empirical basis for research, in which lin-
guistic phenomena are statistically analysed with the help of multivariate
techniques, and in which social and cultural variation is explicitly included
into the multifactorial model. Geeraerts (2006a: vi)
Multivariate statistics is a powerful tool for any social scientist. Quantita-
tive methods are essential in determining the probability that our generali-
sations are representative of the population but also because they help find
patterns of co-occurrence that might be otherwise impossible to identify.
This is especially the case when faced with so complex and multidimen-
sional a phenomenon as language. Without language modules such as se-
mantics, pragmatics, lexis, or syntax, language description must account for
the simultaneous interaction of the different dimensions of language. In-
deed, not only is multifactorial modelling of language, and the various fac-
tors that go into speech, necessary to capture the said complexity, it also
represents a cognitively plausible model of language production.
Multifactorial analysis began by operationalising the various factors that
impact upon the use of lexemes, morphemes, and syntactic patterns. This
was done through feature analysis of large numbers of found examples.
Despite the success of these first attempts (Dirven et al. 1982; Schmid
1993, Geeraerts et al. 1994; Rudzka-Ostyn 1995), manually handling the
results of multifactorial feature analysis limits their interpretative power.
Inspired by the research of psycholinguistics and sociolinguistics, Geer-
aerts et al. (1999), Gries (1999, 2001, 2003), and Speelman et al. (2003)
began applying multivariate statistics to language analysis.
With the development of quantitative tools to treat corpus data and el-
icited data, the future looks bright. It is not difficult to imagine a future
where linguists learn different analytical methods rather than study hypo-
thetical constructs. In such a scenario, with the old dialectic between Cog-
nitive Linguistics and its Mentalist and Structuralist precursors forgotten,
linguists would no longer need to identify themselves as one theoretical
camp or another. Rather, on the basis of empirical and verifiable data, lin-
guists would dispute which method of data collection and analysis is best
suited to answer a given research question. Similarly, they would question
the representativity of data instead of simply choosing a different example
that better suits their hypothesis. Linguists would compare results gleaned
from various methods, and only then, equipped with tested hypotheses,
would they advance theoretical debate. As Geeraerts (this vol.) stresses,
differences in results that either confirm or falsify hypotheses, should be
the only basis for theoretical discussion.
10 Dylan Glynn

3. Quantitative corpus-driven Cognitive Semantics

Employing the usage-based model means that real language use, in all its
complexity, must be the basis of linguistic research. But what does usage-
based methodology entail? From corpus-extracted illustrations of a theo-
retical discussion to collocation studies and then confirmatory multivariate
corpus-driven analysis, there is a full gamut of methodological variation
(cf. Tummers, Heylen & Geeraerts 2005 and Heylen, Tummers & Geer-
aerts 2008 for a discussion on this point; Gilquin & Gries 2009 extend this
discussion to the use of experimental data). Before quantitative corpus-
driven Cognitive Semantics can become a mainstream methodological
choice for linguists, it must answer two very important questions. How can
we use numbers to analyse semantics and why should we do so? The prob-
lem is not corpus data or even corpus-driven analysis. Deignan (2005,
2009), for example, offers excellent examples of non-quantitative corpus-
driven Cognitive Semantic research. The problematic issue is quantifica-
tion. Meaning is an inherently subjective, mentally internal, and non-
observable phenomenon. Why should we attempt to analyse it with quanti-
tative techniques?
Before we can consider the strengths and weaknesses of quantitative
Cognitive Semantics, we must set straight two common misconceptions
about corpus methods that persist within the wider Cognitive Linguistic
community.

3.1. Fallacies about corpus methodology in Cognitive Linguistics

Many linguists believe that corpus-driven research is restricted to the study

of collocations and that quantitative techniques used to treat corpus data are
restricted to obtaining probability values. These two beliefs are erroneous.
Firstly, corpus linguistics is not restricted to the study of collocations and
formal correlations. Although the study of ‘words and the company they
keep’ lies at the origins of corpus linguistics, methods for studying corpus
data have developed much since those times, both in terms of the kinds of
questions that are asked and the kinds of techniques used to answer them.
This is not to say that collocations and syntactic patterns are no longer im-
portant within the field, as we will see in the following chapters, but they
are far from the only research paths possible using corpus techniques.
 Corpus-driven Cognitive Semantics 11

Secondly, quantitative techniques are not just about p-values. The belief
that corpus linguistics is all about probability scores and that corpus lin-
guists do not ‘get their hands dirty’ with careful close language analysis
could not be further from the truth. Moreover, this is particularly false
when corpus techniques are applied to semantic research. Indeed, the
usage-feature or behavioural profile approach, one of the principal trends in
corpus-driven lexical semantic research (Gries 2006a; Divjak 2006; Glynn
2009; Janda & Solovyev 2009; Speelman & Geeraerts 2010) and gram-
matical semantic research (Heylen 2005; Tummers et al. 2005; Wulff et al.
2007; Grondelaers et al. 2008; Szmrecsanyi 2010) involves the detailed and
extremely laborious task of analysing a wide range of formal, semantic, and
sociolinguistic features of thousands of natural language examples. Given
this kind of research, arguing that corpus linguists are just interested in
numbers is risible.
The quantitative treatment of data made available through meticulous
manual or semi-automatic analysis, should not be seen as the hunt for p-
values. Statistics has several roles – estimating statistical signficance; de-
tecting patterns in data; identifying the relative importance of usage factors
in grammatical composition or word choice; and determining whether a
proposed model or explanation accurately describes the data at hand.
Probability values are but one of these roles, they indicate statistical sig-
nificance. They are important because they tell us that our results would
most likely be replicable, but they say no more than that. Save as confir-
mation that a given analysis and its interpretation are more than chance, p-
values offer no linguistic insights per se.
If we accept that corpus linguistics can be more than the study of formal
phenomena and that statistics is more than the search for p-values, then we
can move on to consider the true strengths and weaknesses of quantitative
corpus-driven Cognitive Semantics. These questions are left up to the
authors of this volume, but let us briefly consider the main issues that lie
behind many of the debates in the field.

3.2. Issues for corpus-driven semantics

3.2.1. Corpus representativity

There is, in reality, no such thing as a balanced corpus and no corpus can
ever hope to be representative of a language. Corpus linguists have argued
12 Dylan Glynn

that despite this shortcoming, a corpus is surely more representative than a

single speaker. This response is not as unproblematic as it sounds, raising
fundamental questions about our object of study. Modern linguistics was
founded by de Saussure who delimited the object of study, langue, as a
structured system in a given place and time. His ideas set the stage for the
erroneous assumption that language is a discrete object of study, whose
delimitation is not an issue. In some ways, much of the spilt ink of 20th
century linguistic theory directly follows from disagreement over the object
of study. From syntax and core grammar to pragmatics, linguistics argued
over what language is for most of the preceding century. As stressed above,
Cognitive Linguistics unifies modular language analysis with a holistic
recontextualised approach to language. In light of this, one might argue that
a corpus, which cannot represent the entire complexity of a language, can
never be a basis for studies of language, writ large. The response is straight-
forward - we do not attempt to account for all of language in every study.
The usage-based model places variation, between groups and even be-
tween individuals, as an integral part of language. This we can accept a
priori. The implications of such an assumption are brought forward when
we build our corpora or extract our data for analysis. Indeed, in each cor-
pus-driven study, the linguist is forced to answer the question: what part of
language are we studying? Often practical issues dictate such choices, but
in any case, they are always overtly recognised choices. These choices di-
rectly determine the scope of a study and this information should reflect
variation inherent in language. Nevertheless, fine-tuning our research
methods and building better, more diverse, corpora remain important keys
to improving representativity. Gries (2006b; 2008) examines some of the
issues at hand and the general push towards more analyses of spoken cor-
pora is well documented in corpus linguistic journals (cf. Newman 2008a
and this vol.).
It is impossible to study all of language at once (the reason why the
Structuralist and Modularist programmes chose to focus on ‘parts’ of lan-
guage). Corpus linguistics makes this impossibility overt and, as such, this
apparent limitation is actually a blessing in disguise. It will help linguists be
much more realistic about what can be scientifically said about language in
a single study.
 Corpus-driven Cognitive Semantics 13

3.2.2. Negative evidence

One of the most commonly cited criticisms of corpus linguistics, a criticism

that goes back to Chomsky, is that there is no negative evidence. No cor-
pus, irrespective of its size, can possibly represent a language, let alone tell
us whether a given expression or use of an expression is impossible. How-
ever, negative evidence in this sense of the term is only crucial in a rule-
driven theory of language, such as that propounded by Chomsky. Since, in
a usage-based model, grammaticality is based on entrenchment, and en-
trenchment is a result of use varying from individual to individual (indeed,
even within the individual), there are no hard and fast rules. In a usage-
based model of grammar, grammatical rules are merely generalisations
about usage. For such a model of language, negative evidence is of much
less importance.
To take a simple example, in a Mentalist rule-driven grammar, for the
proposal of a rule that the second person copula in English takes the form
are, any occurrence of *you is or *you am would disprove the rule. How-
ever, since grammaticality is a matter of degree in Cognitive Linguistics,
the fact that in certain regions of England certain social groups say you is
(especially in the reduced form) is incorporated into our grammar, without
negating the constructional importance of the you are pattern. Instead, the
use of this construction is relative to established sociolinguistic factors and
is part of the system. Hence, grammatical tests and their need for negative
evidence are of no concern to the usage-based linguist. Langacker’s oper-
ationalisation of grammaticality means that frequent is grammatical and
infrequent is less grammatical.
For Structuralism, defining the meaning of a lexeme, such as bachelor,
might be an example of the need for negative evidence. If we define a
bachelor as an unmarried man, but there exists a propositional use that
contradicts this meaning, such as ??a Catholic priest is a bachelor, the
definition would be shown to be inadequate. In this case, one would need to
redefine the lexeme (here as a “man thought of as someone who could
marry” Wierzbicka 1990). The process is methodologically identical to the
rule (re-)writing of Generative Grammar. If a proposed rule generates sen-
tences that are unacceptable, then the rule is too powerful and needs to be
redefined. In this way, the desire to identify the necessary and sufficient
conditions for distinguishing the meaning of one lexeme from the meaning
of another motivates the need for negative evidence. However, the idea of a
proof, which may be falsified with a single counter example, is arguably
14 Dylan Glynn

not informative for language description. Rather, social sciences are better
served by generalisations based on samples. Given the model of language,
the need for negative evidence is not so obvious for the cognitive linguist.
For further discussion on this issue, within the framework of corpus Cogni-
tive Linguistics, see Stefanowitsch (2006, 2008).

3.2.3. Frequency and salience

What can frequency tell us about language? This is a very real issue for the
corpus-driven study of meaning. Although some might argue that fre-
quency is an indicator of productivity or entrenchment, it would be difficult
6
to claim that frequency directly equates salience. If salience is an indicator
of relative semantic importance, we are faced with a problem. In terms of
semantic content, most frequent often equates least semantically important,
where rarity, or marked usage, indicates greater semantic importance.
Givón (1991) is explicit on this point - “The marked category [...] tends to
be less frequent, thus cognitively more salient, than the corresponding
unmarked one.” However, in another sense of semantic importance, the
reverse is sometimes true. For example, when a language more frequently
refers to a given phenomenon, cultural or otherwise, we assume that it is
culturally significant. This line of reasoning is basic to much of Wierzbicka
(1985) and Lakoff’s (1987) research on culturally determined categories.
Salience is often cited as a crucial notion for establishing relations be-
tween categories, especially prototype structures (Durkin & Manning 1989,
Geeraerts 2000). It follows that since the results of psycholinguistics are
often based on relative categorisation determined by salience, understand-
ing the relationship between frequency and salience is essential for the
comparison of corpus and experimental results. Thus, although there is a
relationship between frequency, salience, and semantic import, this rela-
tionship is not straightforward. Since corpus research is dependent on the
study of relative frequency, how can this method be used to talk about se-
mantic structure and how can its results be compared to experimental re-
sults?
This question may be answered by operationalising the concepts in-
volved. It appears that there are different kinds of salience, yet how fre-
quency is related to perceptual salience, conceptual cultural salience, or
formal linguistic salience is still an open question. This debate is growing
within the field and is treated by Newman (this vol.) and Schmid (this vol.),
 Corpus-driven Cognitive Semantics 15

but see also Gilquin (2006, 2010), Wiechmann (2008a, 2008b) Gilquin &
Gries (2009), and Arppe et al. (2010). The answers to such questions will
probably be found when concepts such as salience are more clearly defined
7
through operationalisation. Perhaps, it will be shown that frequency can be
a measure of certain kinds of salience and not others. Currently, however,
the debate remains open.

3.3. Advantages of corpus-driven Cognitive Semantics

This volume should not only be seen as an attempt at identifying the limita-
tions of a corpus-driven method, the advantages must also be developed.
Let us briefly consider the basic advantages of this approach.

3.3.1. Empirical

It may seem obvious that corpus-driven research is empirical, but it is im-

portant to state this clearly because, for two different reasons, many within
the research community believe that the use of corpus data does not consti-
tute empirical research in linguistics. The first argument comes from within
the experimental community and is discussed by Gries & Divjak (this vol.).
Summarised briefly, the argument is that as corpus linguists, we do not
have direct access to the mind, and therefore language production. It fol-
lows that we are not actually studying language production, but traces of
language use extant in corpora. The response to this critique is simple - it is
true. No corpus linguist studies language production. However, these traces
of use are a wonderfully rich source of information on how language is
used. Moreover, although we have no direct access to our object of study,
nor does experimentation. Psycholinguists elicit responses to stimuli; they
do not look inside the ‘black box’ itself. Thus, we collect data, they elicit it.
Neither approach has direct access to the mind or its functioning.
Psycholinguistic experimentation is better placed to look at processing and
dynamic issues in language production, while corpus linguistics is better
placed to look at natural use and, perhaps, even the intersubjective dimen-
sion of language (cf. Glynn & Krawczak submitted). Although, as stressed
by Newman (2008a; this vol.), there is a terrible lack of spoken and interac-
tive corpora for exploring these dimensions of language.
16 Dylan Glynn

The second argument comes from the community that uses introspection
to perform linguistic analysis. Their argument is reasonable - since our
theory of language is a semantics-driven theory of language and semantics
is necessarily subjective, that is internal to the mind, introspection is the
only viable method for its analysis. Talmy (2008: xix) is explicit about this
point. He stresses that corpus research “cannot directly yield many abstract
linguistic patterns”. His argument is similar to that of the psycholinguists’,
save, at least, that he is correct in saying that introspection does access the
mind (unlike in psycholinguistic experimentation). His point is not disputed
but we have two important additions to the discussion. Firstly, despite this
limitation, the patterns of natural language usage are an incredibly rich
source for working out how people use language and this can produce a
very good picture of language structure. Secondly, as stressed by Geeraerts
(this vol.), introspection is vital to corpus linguistics. The categories chosen
for study, the actual analysis of those categories, and, of course, the hy-
potheses that the study will test are all a result of introspection. It is not that
empirical research replaces introspection, rather introspection is used to
propose hypotheses, which then need to be tested by operationalising the
questions and designing a study that will adequately answer those ques-
tions. If we cannot find a means to adequately answer a scientific question,
then we are in the same position as scientists in all fields and must continue
to look for a means to do so. This is no argument to remain at the hypo-
thetical stage of enquiry, based entirely upon introspection.

3.3.2. Quantitative

If one has empirical data, then one may quantify them. However, it is far
from self-evident that corpus-driven semantic research should be quantita-
tive. Since debating this proposal is the point of the current volume, we will
not delve too deeply into this discussion. However, the basic advantage
offered by a quantitative approach to semantics is that it permits the oper-
ationalisation of our studies. The benefits of, and indeed need for, oper-
ationalisation are discussed by Stefanowitsch (this vol.). The questions, of
course, are – can and should we operationalise semantics quantitatively?
These questions are specifically broached by Newman (this vol.) and
Schmid (this vol.) but, to a greater or lesser extent, they are considered by
all the contributions to this volume. One basic point should be established -
no one wishes to argue that we should reduce semantic questions to num-
 Corpus-driven Cognitive Semantics 17

bers. It is not the frequency per se of linguistic features that is of interest,

but what this says about usage, the relative association of forms and mean-
ings in context.
Even if quantification is possible, the next question is - why do we need
it, what do we gain from employing quantitative techniques? Firstly, they
permit, and indeed encourage, the empirical cycle, explained by Geer-
aerts (this vol.). It is much easier for the research community to verify the
results of a study when one can see exactly how those results were ob-
tained. Secondly, quantified data can be examined using statistics. Statisti-
cal analysis of data offers several advantages: (i) Confirmatory statistics
allows one to determine the statistical significance of the results of an an-
alysis. In other words, what is the probability that similar results would be
obtained if further examples were examined in the same way? (ii) Multi-
variate statistics allows one to identify patterns in usage that would be ef-
fectively impossible to identify using introspection. (iii) Statistical model-
ling allows one to test the accuracy of an analysis. This kind of measure
determines how much of the variation, in a given sample, a given analysis
can explain. These multivariate techniques do not only examine the effects
of numerous factors of usage simultaneously, they include the interaction of
these factors as well. Thus, Empiricism permits quantification and quantifi-
cation permits statistical confirmation but also multifactorial analysis. This
is the third basic advantage.

3.3.3. Multifactorial

Empirical research permits quantification, and, in turn, quantification per-

mits multifactorial analysis. Although this may seem like a narrow addition
to the discussion, it is arguably the most important facet of corpus-driven
semantic research. As stressed in section 2.3, it is the complexity of our
object of study that is our greatest hurdle. The cognitive model of language
insists that we simultaneously take into account the full socio-cognitive
spectrum of factors that influence and motivate language. Whether the re-
search is based on questionnaires, advanced eye-tracking experiments,
word space models derived from mass computation of huge corpora, or the
nitty-gritty manual analysis of semantic features of a small corpus sample,
multivariate modelling is a basic and crucial tool. It is almost impossible to
imagine how one might account for all the factors, and the interaction of
these factors, that go into language production and comprehension. How-
18 Dylan Glynn

ever, should it ever be possible, it will only be so with the aid of such ana-
lytical tools. Therefore, arguably, it is precisely the possibility of multifac-
torial analysis that makes quantitative empirical research, be that corpus-
driven or experimental, essential for linguistic semantics.

4. Cognitive Semantics – lexical and grammatical meaning

This decade has seen a veritable surge of corpus-driven Cognitive Lin-

guistic research. This volume can only scratch the surface of the theoretical
questions that such a method raises and offer a small sample of the research
it makes possible. The focus of the volume is specifically the application of
quantitative corpus-driven methods to Cognitive Semantics. Yet, even lim-
iting the horizons in this manner, the possible applications and issues go
beyond the purview of any single volume. Let us place the work presented
here in its methodological context. This will allow the reader to better
understand its importance, but also see where future directions of the field
lie.

4.1. Survey of the field

4.1.1. Semasiology–onomasiology and semantic schematicity

For Cognitive Semantics, one has two possible objects of study. Firstly, one
may ask what forms are available to express a given concept (onomasiol-
ogy or synonymy). Secondly, one can ask what concepts are expressed by a
given form (semasiology or polysemy). Having categorised the research as
onomasiological or semasiological, one may further divide it into the study
of schematic or non-schematic form and meaning. Schematic meanings are
more typically expressed by grammatical forms, such as morphemes and
grammatical constructions while non-schematic meanings are more typi-
cally expressed by lexical forms. The semasiological–onomasiological
division crosses both objects of study. For grammatical meaning, the ono-
masiological study of schematic concepts includes, for example, syntactic
alternations or the choice between grammatical cases. The semasiological
study of grammatical forms looks more closely at a given construction or
morpheme. The same distinction applies to lexical semantics, effectively
drawing a distinction between the study of polysemy networks and near-
 Corpus-driven Cognitive Semantics 19

synonymy lexical fields. This leads to four research questions – (i) What
words do people choose? (ii) What grammatical constructions do people
choose? (iii) How do people use a word? (iv) How do people use a gram-
matical construction? We can label these research questions respectively as
lexical synonymy, grammatical synonymy, lexical polysemy, and gram-
matical polysemy.
Although there is no clear divide between lexical and grammatical
forms and obviously even less so between the schematic and non-schematic
meaning, the onomasiological–semasiological division is reasonably clear-
cut. This ‘form-first’ or ‘concept-first’ distinction is arguably fundamental
to semantic research (Geeraerts et al. 1994; Geeraerts 2006a). However,
even here, Glynn (2010) argues that there exists a continuum between these
two dimensions, depending on the granularity of the study. This is because
all variation in prosody, syntax, and morphology is technically a change in
form that can represent a subtle change in meaning. To this extent, the line
between semasiology and onomasiology is also not ultimately discernable.
Newman (2008b; this vol.) broaches this question in his discussion on the
problems of studying the lemma, arguably a very coarse-grained level of
analysis. Nevertheless, a methodological division does exist, which de-
mands that research begin with either a concept – examining the various
forms that express it, or with a form – examining the various meanings
expressed by it.
The four research domains listed above can be used to categorise all re-
search in Cognitive Semantics. The work on culturally determined con-
cepts, semantic frames, cognitive models, and metaphors (Wierzbicka
1985, Kövecses 1986, Lakoff 1987: case study 1) is obviously lexical ono-
masiology, just as the research in polysemy is typically lexical semasiology
(Lakoff 1987: case study 2, Fillmore & Atkins 1992, Cuyckens 1995). The
studies in syntactic alternations, popular in Construction Grammar, are
clearly grammatical onomasiology. Similarly, Langacker (2000) and Talmy
(2000) examine the differences between various grammatical profilings, an
inherently schematic-semantic, morpho-syntactic, and onomasiological
domain of research. This leaves the semasiological study of individual
grammatical forms and syntactic patterns. Such research is less well-known
but equally represented. Important studies include Lakoff (1987: case study
3), Janda (1993), Goldberg (1995), Rudzka-Ostyn (1996), Dąbrowska
(1997), Geeraerts (1998).
Having established the four research domains of Cognitive Semantics,
we need to introduce two methodological variables specific to corpus-
20 Dylan Glynn

driven analysis - the relative use of observable objective linguistic features

and the relative use of statistical techniques.

4.1.2. Hard statistics–soft statistics and objectivity

Methodologically, corpus-driven research in semantics is far from being a

single established approach. There are different ways of collecting data,
analysing the data, and then a wide range of quantitative techniques avail-
able for treating the results of these analyses.
First, research in a quantitative semantic study can be restricted to the
analysis of formal observable phenomena or it can include semantic non-
observable phenomena. This is important since the inclusion of non-
observable, and therefore subjectively determined, factors in semantic an-
alysis greatly reduces the objectivity of the study. Moreover, if the analysis
is restricted to formal observable characteristics, one has the possibility of
the automatic treatment of data. It follows that if the analysis is automatic,
one may examine large quantities of examples. As corpora increase in size,
and as methods for automatically treating them improve, new possibilities
for this line of research are beginning to emerge. Examples of automatic
analyses that seek to answer semantic questions include the collostructional
analyses of Stefanowitsch & Gries (2003), Gries & Stefanowitsch (2004)
and Hilpert (2008) and the lexical onomasiological studies of Arppe (2008)
and Levshina et al. (forthc.). However, the application of computational
techniques, such as word space modelling, to Cognitive Semantic research
is also emerging. The Sem•metrix project, led by D. Geeraerts, is an exam-
ple of such an approach (Peirsman et al. 2010).
The lack of semantic richness in formal automatic analysis is offset by
the fact that its results can be more reliable, not only due to the objective
nature of the analysis, but because a larger sample improves representa-
tivity. Nevertheless, despite their merit, such approaches arguably miss
much of what is important for semantic research. It is questionable whether
linguists will ever be able to sufficiently describe language structure based
exclusively upon formal patterns of usage. In light of this, small-scale de-
tailed manual semantic analysis, typical of the usage-feature behavioural
profile approach, remains an important line of investigation. Both lines of
research, as well as their pros and cons, are treated in this volume.
Determining the degree of statistical sophistication for the treatment of
data is not a straightforward question. With small quantities of data, due to
 Corpus-driven Cognitive Semantics 21

limited resources imposed by transcribing spoken language, scarcity of

historical sources, or the practical constraints of fieldwork, one can do little
more than count and compare raw numbers. Moreover, the meticulous
manual analysis of usage features considerably restricts the number of ex-
amples in a dataset, in turn, restricting the efficacy of more advanced statis-
tical techniques. Finally, the technical competence of the author is a real, if
not scientific, factor that influences the choice of statistical methods. At
least for the current generations of linguists, statistical training is rare,
though this is set to change in the near future.
Quantitative approaches to semantic structure within Cognitive Lin-
guistics begin with counting examples of a certain kind and comparing
them to examples of another kind (e.g.: Hanegreefs 2004; Davidse et al.
2008; Dziwirek & Lewandowska-Tomaszczyk 2009). Such quantitative
data may employ tests for statistical significance, such as the t-Test or Chi-
square test. Other quantitative research employs exploratory techniques
such as Cluster Analysis and Correspondence Analysis (Gries 2006a,
Divjak 2006, Szelid & Geeraerts 2008, Glynn 2009). These techniques may
also be combined with significance testing, but their focus is identifying
patterns and associations in the data. Finally, confirmatory statistics such as
Linear Discriminant Analysis and Logistic Regression are complex, yet
extremely powerful techniques (Gries 2003, Heylen 2005, Speelman and
Geeraerts 2010). Confirmatory statistical modelling presents many possi-
bilities for the multivariate description of language, as well as the verifica-
tion of that description. Glynn & Robinson (in press) offers a survey of the
statistical approaches popular in Cognitive Semantics.

4.1.3. Social dimensions of semantics

Special mention must be made of the sociolinguistic element in corpus-

driven research. This line of research is of growing importance (Geeraerts
2005, Croft 2009, Geeraerts & Speelman 2010) and is especially significant
to quantitative corpus-driven approaches for two reasons. Firstly, sociolin-
guistics, along with psycholinguistics, has employed some of the most ad-
vanced statistics for language study and has, to some extent, paved the way
forward for Cognitive Semantics. Secondly, it is precisely this element of
language structure that is effectively impossible to account for using intro-
spection and difficult to adequately account for using experimentation. It is
for these reasons that the interaction of meaning and society is at the heart
22 Dylan Glynn

of corpus-driven Cognitive Semantic research. Two recent anthologies,

Kristiansen & Dirven (2008), Geeraerts et al. (2010), are devoted to the
subject.

4.2. Case studies in the field

We can now consider the field of research. Tables 1 and 2 were compiled
by collecting the relevant studies from a range of sources. These sources
include the anthologies listed in footnote 2, four Cognitive Linguistics
journals (Cognitive Linguistics, Annual Review of Cognitive Linguistics,
Language and Cognition and Constructions and Frames), and three im-
portant corpus linguistics journals (Corpus Linguistics and Linguistic
8
Theory, Corpora, International Journal of Corpus Linguistics). Tables 1
and 2 list both the object of study and the method employed. The method-
ological information offered includes the degree of statistical complexity:
counts, collocation (including Collostructional Analysis), exploratory stat-
istics (Hierarchical Cluster Analysis HCA; Principal Component Analysis
PCA; Configural Frequency Analysis CFA; Multiple Correspondence An-
alysis MCA; Multidimensional Scaling MDS), and confirmatory multifac-
torial analysis (Linear Discriminant Analysis LDA; Profile-Based Analysis
PBA; Logistic Regression Analysis LRA; Mixed Effects Logistic Regres-
sion MER). These techniques are explained by various authors in Glynn
9
and Robinson (in press). When more than one technique is employed, only
the most advanced is listed. The degree of objectivity is also indicated –
studies based entirely on formal analysis of observable data versus studies
that include at least some non-observable semantic usage-features. The
tables also indicate if a study includes extralinguistic dimensions – socio-
linguistic, diachronic, or contrastive factors. The studies are sorted chrono-
logically, relative to whether they have a semasiological or onomasiologi-
cal emphasis and whether they treat lexical or grammatical forms and
concepts.
 Corpus-driven Cognitive Semantics 23

Table 1. Corpus-Driven Cognitive Semantics - Lexical Synonymy and Polysemy

Object Method Reference
Lexical Onomasiological
Verbs of informing (Eng.) - counts sem Dirven et al. 1982
Nouns for clothing (Dch.) socio. counts sem Geeraerts 1993b; et al. 1994
Verbs of starting (Eng.) - counts sem Schmid 1993
Verbs of response (Eng.) diach. counts sem Rudzka-Ostyn 1995
Terms for football (Dch.) socio. conf. PBA sem Geeraerts et al. 1999
Abstract nouns (Eng.) - counts sem Schmid 2000
Verbs of posture (Eng.) - counts form Newman & Rice 2004a
Concept of ANGER (Eng.) diach. counts sem Gevaert 2005
Verbs of eating (Eng.) - counts form Newman & Rice 2006
Verbs of intention (Rus.) - explor. HCA sem Divjak 2006
Verbs of attempting (Rus.) - explor. HCA sem Divjak & Gries 2006
Verbs of becoming (Spanish) - explor. MDS sem Bybee & Eddington 2006
Verbs of separation contr. explor. HCA sem Majd & Bowerman 2007
(Dch., Eng., Germ., Swed.)
Verbs of cognition (Fin.) - conf. LRA form Arppe 2009
Concepts POSITIVE, NEGATIVE socio. explor. MCA sem Szelid & Geeraerts 2008
(Hungarian)
Verbs of starting (Eng., Rus.) contr. explor. HCA sem Divjak & Gries 2009
Verbs of reception (Dch.) diach. counts form Delorge 2009
Adverbs for ‘again’ (Mnd.) - conf. CFA sem Jing-Schmidt & Gries 2009
Concepts HAPPY, SAD (Rus.) - explor. HCA sem Janda & Solovyev 2009; in press
Concepts LOVE, HATE (Pol.) contr. colloc. count form Dziwirek & Lewandowska 2009
Verbs of posture (Eng.) diach. colloc. count form Newman 2009
Terms for BOTHER (Eng.) socio. conf. LRA sem Glynn 2010
Terms for FOOTBALL (Port.) socio. conf. PBA sem Soares da Silva 2010
Verbs of causation (Dch.) socio. conf. LRA sem Speelman & Geeraerts 2010
Verbs come, go (Jap., Kor.) contr. counts sem Kabata & Lee 2010
Verbs of bothering (Eng.) socio. conf. MER sem Glynn in press a
Verbs of cognition (Pol.) socio. conf. LRA sem Fabiszak et al. in press
Verbs of causation (Dch.) socio. conf. MER form Levshina et al. in press
Verbs of possibility (Eng., Fr.) contr. conf. LRA sem Deshors & Gries in press
Concept LIBERTY (Eng.) socio. conf. LRA sem Glynn forthc.
Lexical Semasiological
Verb ask (Eng.) diach. counts sem Rudzka-Ostyn 1989
Verb need to (Eng.) socio. colloc. count sem Nokkonen 2006
Verb run (Eng.) - explor. HCA sem Gries 2006a
Verb, Noun hassle (Eng.) socio. explor. MCA sem Glynn 2009
Verb, Noun bother (Eng.) socio. conf. LRA sem Glynn this volume
Adjective deep (Germ.) - colloc. sem Zeschel this volume
Verb annoy (Eng.) socio. explor. MCA sem Glynn submitted
Legend: Adj. - Adjective; Adv. - Adverb; Alt. - Alternation; Cx. - Construction; Imperf. - Imperfective;
Perf. - Perfective; Dch. - Dutch; Eng. - English; Fin. - Finnish; Fr. - French; Germ. - German; Jap. -
Japenese; Kor. - Korean; Mnd. - Mandarin; Pol. - Polish; Port. - Portugese; Rus. - Russian; Swed. -
Swedish; Ukr. - Ukrainian; contr. - contrastive; diach. - diachronic; socio. - sociolinguistic; confirm. -
confirmatory statistics; explor. - exploratory statistics; sem - semantic annotation, form - formal annota-
tion.
24 Dylan Glynn

Table 2. Corpus-Driven Cognitive Semantics – Grammatical Synonymy and Polysemy

Object Method Reference
Grammatical Onomasiological
Alt. Phrasal Verb syntax (Eng.) - counts form Gries 1999
Alt. Future Cxs (Eng.) - counts form Szmrecsanyi 2003
Alt. Adj. Syntax (Eng.) - confirm. LDA sem Wulff 2003
Alt. Adj. Inflection (Eng.) - explor. PCA form Gries 2003
Aspect Prepositions (Eng.) - counts form Newman & Rice 2004b
Alt. Adj. Inflection (Dch.) socio. confirm. LRA form Tummers et al. 2005
Alt. Middle Field Syntax (Germ.) socio. confirm. LRA sem Heylen 2005
Alt. go V–go and V Cxs (Eng.) - colloc. collostr. form Wulff 2006
Grammatical Cases in Slavic contr. explor. MDA sem Clancy 2006
Alt. Dative socio. confirm. LRA sem Bresnan et al. 2007
Alt. Presentative Cxs (Dch.) socio. confirm. LRA sem Grondelaers et al. 2007; 2008
Temp. Adv. Clause Syntax (Eng.) - confirm. LRA sem Diessel 2008
Alt. Nominal – Clausal Compls (Eng.) - colloc. collostr. form Wiechmann 2008b
Alt. Act-Pass. Voice; Alt. Future Cxs; socio. confirm. CFA form Stefanowitsch & Gries 2008
Alt. Phrasal Verb syntax (Eng.)
Alt. Dep. Clause syntax (Dch.) socio. confirm. LRA sem De Sutter 2009; et al. 2008
Alt. Adj. Inflection (Dch.) socio. colloc. collostr. form Speelman et al. 2009
Alt. Genitive Cxs (Eng.) socio. confirm. LRA sem Szmrecsanyi 2010
Alt. Imperf.–Perf. Cxs (Rus.) - confirm. MER form Divjak this vol.
Alt. Case – Adposition (Estonian) socio. confirm. LRA sem Klavan in press
Grammatical Semasiological
Case Dative (Pol.) - counts sem Rudzka-Ostyn 1996
Cxs. think nothing of Gerund; Imper.; - colloc. collostr. form Stefanowitsch & Gries 2003
Progr.; Ditrans. (Eng.)
Cx. into-Causative (Eng.) - confirm. CFA form Gries & Stefanowitsch 2004
Cx. into-Causative (Eng.) socio. colloc. collostr. form Wulff et al. 2007
Cxs Future diach. colloc. collostr. form Hilpert 2008a
(Eng., Dan., Swed., Germ.) contr.
Cx. Comparative (Eng.) - colloc. collostr. sem Hilpert 2008b
Aux. shall; Aspect Perfect (Eng.) diach. explor. HCA form Gries & Hilpert 2008
Cx. good PrP (Eng.) - counts sem Zeschel 2009
Cx. mit-Predicative (Germ.) - confirm. CFA form Hilpert 2009
Cx. Dative (Dch.) diach. colloc. collostr. form Colleman 2009, this vol.
Cx. Raising (Eng., Dch.) diach. colloc. collostr. form Noël & Colleman 2010
contr.
Cx. Verb Poss. way (Eng.) - explor. HCA form Gries & Stefanowitsch 2010
Cx. Benefactive Ditransitive (Eng.) socio. counts form Colleman 2010
Cx. V – NP (Eng.) - colloc. collostr. form Wulff this vol.
Cx. Gerund Compl. Clause (Eng.) - colloc. collostr. form Hilpert this vol.
Cx. for Durative Adv. (Eng.) - colloc. collostr. form Fuhs this vol.
Suffix dissposseive –ont (Dch.) diach counts sem Delorge & Colleman in press
Cx. Epistemic stance (Eng.) socio confirm. MER sem Glynn & Krawczak submitted
 Corpus-driven Cognitive Semantics 25

The first thing one will notice looking at the tables is that for each object
of study, there is a tendency towards more sophisticated quantitative tech-
niques over time. Just as noticeable is the tendency for semantic analysis in
lexical research contrasted by formal analysis in grammatical research.
Although this is in part due to a large amount of collocation studies of
grammatical constructions, there is no reason to suppose that collocation
should not be equally applied to the study of lexical semantics. As
Zeschel (this vol.) shows, the collocation tradition, which was originally
lexically orientated, still has a great deal to offer the study of lexical seman-
tics. Noteworthy as well is a small increase in the amount of studies includ-
ing sociolinguistic, diachronic, and contrastive parameters in their research.
The tables also reveal the biases in research interests across the research
paradigm. In lexical research, verbs obviously have an over representation,
nouns, adjectives, and adverbs being poorly represented. Also, and some-
what surprisingly, seeing the history of Cognitive Semantics, prepositions
are entirely absent. Some of this bias is because verbs, due to their argu-
ment structure and its participants, offer more observable features for an-
alysis, facilitating quantitative study. Overcoming the difficulties of study-
ing other parts of speech should, nevertheless, be a goal in the research
field. Equally noticeable is the emphasis upon onomasiological research.
Again, there is a likely methodological reason. It is relatively easy to dis-
tinguish the use of forms compared to distinguishing uses, especially since
formal and extralinguistic factors may often contribute a lot to explaining
onomasiological variation. The research field needs to approach the diffi-
culties of semasiological lexical description and this volume offers two
contributions, Zeschel (this vol.) and Glynn (this vol.), that attempt this.
The grammatical research also shows clear biases. Obviously, the popu-
larity of the Collostructional Analysis accounts for much of this. However,
the fact that effectively all onomasiological research examines syntactic
alternations and all semasiological research looks at syntactic-lexical pair-
ing is a clear and serious shortcoming. Although there are some important
exceptions to this trend, such as the work on adjectival inflections and as-
pectual categories, the field must move towards other kinds of grammatical
semantics. The semantics encoded by phonological structures, such a
prosody, are entirely absent. The lack of research in grammatical case is
especially surprising, seeing the rich tradition of this in Cognitive Linguist-
ics. This volume corrects this imbalance with Divjak’s (this vol.) research
on aspect. However, both the semasiological and onomasiological research
in grammatical semantics needs to broaden its field of research.
26 Dylan Glynn

Perhaps, the most striking gap in the research is the lack of work on dis-
course questions and broad culturally-determined concepts, such as those
proposed by Lakoff (1987). The first steps have been taken in the study of
Idealised Cognitive Models (Dziwirek & Lewandowska 2009, Glynn
forthc.), but this volume does not offer any such research. Extending this
method to more culturally-orientated concepts is a clear goal for future
research. The study of discourse is difficult due to the lack of adequate
spoken corpora. Fischer (this vol.) offers an important contribution on this
front by developing a corpus especially designed to capture such discourse
structures. As the field matures, such lines of inquiry will certainly attract
more attention. The next step is to develop the methods to answer broader
semantic research questions. This volume makes a small, but important
step, in this direction.

Notes

1. I would like to thank G. Gilquin, M. Hilpert, and C. Paradis for their help. All
shortcomings remain my own.
2. Testimony to the surge in corpus-driven research is the large number of an-
thologies, devoted entirely or substantially, to corpus-based Cognitive Lin-
guistics. A list of these includes Gries & Stefanowitsch (2006), Stefanowitsch
& Gries (2006), Gonzalez-Marquez et al. (2007), Kristiansen & Dirven
(2008), Zeschel (2008), Lewandowska-Tomaszczyk & Dziwirek (2009),
Gilquin (2009), Gries et al. (2009). Geeraerts et al. (2010), Rice & Newman
(2010), Marzo et al. (2010), Glynn & Robinson (in press), and Gries & Divjak
(forthc.). Geeraerts (2006b) offers a detailed description of the generalisation
of this trend across the field.
3. R. Gibbs and M. Tomasello lead the field. One cannot even begin to list the
most important contributions here, but early studies include Gibbs (1990) and
Tomasello (2000).
4. Although one may interpret Langacker’s position in this way, from a strictly
cognitive perspective, the idea of competence is not at all relevant. Cf. Paradis
(2003) for a discussion on the notion of competence.
5. The use of frequency as a means for operationalising grammar deserves dis-
cussion beyond the scope of this introduction. Stefanowitsch (this vol.) shows
its importance at a methodological level and Bybee (2007), amongst others,
has fully integrated this notion into her research.
 Corpus-driven Cognitive Semantics 27

6. Langacker (1987: 100) is explicit about the relationship between entrench-

ment and frequency. Schmid (2000: 39) is still more explicit, proposing, what
he terms, the Corpus-to-Cognition Principle – “Frequency in text instantiates
entrenchment in the cognitive system.” Schmid has since, however, distanced
himself from this position.
7. Gries et al. (2005, 2010), Arppe & Järvikivi (2007) and Wulff (2009b) exam-
ine this question in empirical terms. Jones et al. (2007) also offer an interest-
ing means of operationalising these kinds of issues. Geeraerts (2000) directly
addresses the issue of salience in lexical semantics.
8. For sake of brevity, recent monographs in the field have not been included.
For the reader’s reference, books devoted to the field include Geeraerts et al.
(1999), Schmid (2000), Fischer (2000), Gries (2003), Mukherjee (2005),
Szmrecsanyi (2006), Hilpert (2008), Arppe (2008), Wulff (2009a), Divjak
(2010), and Gilquin (2010).
9. Specific references for these techniques include: Collostructional Analysis -
Hilpert (in press), Stefanowitsch & Gries (2003), Gries & Stefanowitsch
(2004), Hierarchical Cluster Analysis - Divjak & Fieller (in press), Logistic
Regression Analysis - Baayen (2008) and Speelman (in press), Mixed Effects
Regression - Baayen (in press), Multiple Correspondence Analysis - Glynn (in
press b).

References

Arppe, Antti
2008 Univariate, Bivariate and Multivariate Methods in Corpus-Based
Lexicography – A study of synonymy. Helsinki: University of Hel-
sinki.
2009 Linguistic choices vs. probabilities - how much and what can lin-
guistic theory explain? In The Fruits of Empirical Linguistics. Vol.
1. Process, Sam Featherston and Susanne Winkler (eds.), 1-14. Ber-
lin / New York: Mouton de Gruyter.
Arppe, Antti and Juhani Järvikivi
2007 Every method counts: Combining corpus-based and experimental
evidence in the study of synonymy. Corpus Linguistics and Lin-
guistic Theory 3: 131–159.
Arppe, Antti, Gaëtanelle Gilquin, Dylan Glynn, Martin Hilpert, Arne Zeschel
2010 Cognitive corpus linguistics: Five points of debate on current theory
and methodology. Corpora 5: 1-27.
28 Dylan Glynn

Baayen, R. Harald
2008 Analyzing Linguistic Data: A practical introduction to statistics
using R. Cambridge: Cambridge University Press.
in press Multinomial mixed models. In Polysemy and Synonymy. Corpus
methods and applications in Cognitive Linguistics, Dylan Glynn and
Justyna Robinson (eds.). Amsterdam / Philadelphia: John Benja-
mins.
Bresnan, Joan, Anna Cueni, Tatiana Nikitina and Harald Baayen
2007 Predicting the Dative alternation. In Cognitive Foundations of Inter-
pretation, Gosse Bouma, Irene Kraemer and Joost Zwarts (eds.), 69–
94. Amsterdam: Royal Netherlands Academy of Science.
Bybee, Joan
2007 Frequency of Use and the Organization of Language. Oxford: Ox-
ford University Press.
Bybee, Joan and David Eddington
2006 A usage-based approach to Spanish verbs of becoming. Language
82: 323-354.
Clancy, Steven J.
2006. The topology of Slavic case: Semantic maps and multidimensional
scaling. Glossos 7: 1-28.
Colleman, Timothy
2009 The semantic range of the Dutch double object construction: A col-
lostructional perspective. Constructions and Frames 1: 190-221.
2010 Lectal variation in constructional semantics: “Benefactive” ditransi-
tives in Dutch. In Advances in Cognitive Sociolinguistics, Dirk
Geeraerts, Gitte Kristiansen and Yves Peirsman (eds.), 191-221.
Berlin / New York: Mouton de Gruyter.
Croft, William
2009 Toward a social Cognitive Linguistics. In New Directions in Cogni-
tive Linguistics, Vyvyan Evans and Stéphanie Pourcel (eds.), 395-
420. Amsterdam / Philadelphia: John Benjamins.
Cuyckens, Hubert
1995 Family resemblance in the Dutch spatial prepositions door and
langs. Cognitive Linguistics 6: 183-207.
Dąbrowska, Ewa
1997 Cognitive Semantics and the Polish Dative. Berlin / New York:
Mouton de Gruyter.
Davidse, Kristin, Lieselotte Brems and Liesbeth De Smedt
2008 Type noun uses in the English NP: A case of right to left layering.
International Journal of Corpus Linguistics 13: 139–168.
 Corpus-driven Cognitive Semantics 29

De Sutter, Gert
2009 Towards a multivariate model of grammar: The case of word order
variation in Dutch clause final verb clusters. In Describing and
Modeling Variation in Grammar, Andreas Dufter, Jürg Fleischer
and Guido Seiler (eds.), 225-254. Berlin / New York: Mouton de
Gruyter.
De Sutter, Gert, Dirk Speelman and Dirk Geeraerts
2008 Prosodic and syntactic-pragmatic mechanisms of grammatical vari-
ation: The impact of a postverbal constituent on the word order in
Dutch clause final verb clusters. International Journal of Corpus
Linguistics 13: 194-224.
Deignan, Alice
2005 Metaphor and Corpus Linguistics. Amsterdam / Philadelphia: John
Benjamins.
2009 Searching for metaphorical patterns in corpora. In Contemporary
Corpus Linguistics, Paul Baker (ed.), 9-31. London: Continuum.
Delorge, Martine
2009 A diachronic corpus study of the constructional behaviours of recep-
tion verbs in Dutch. In Studies in Cognitive Corpus Linguistics, Bar-
bara Lewandowska-Tomaszczyk and Katarzyna Dziwirek (eds.),
249 -272. Frankfurt/Main: Peter Lang.
Delorge, Martine and Timothy Colleman
in press The polysemy of Dutch ‘transfer’ constructions: The case of ont-
verbs. In Polysemy and Synonymy. Corpus methods and applications
in Cognitive Linguistics, Dylan Glynn and Justyna Robinson (eds.).
Amsterdam / Philadelphia: John Benjamins.
Deshors, Sandra and Stefan Th. Gries
in press A case for the multifactorial assessment of learner language: The
uses of may and can in French-English interlanguage. In Polysemy
and Synonymy. Corpus methods and applications in Cognitive Lin-
guistics, Dylan Glynn and Justyna Robinson (eds.). Amsterdam /
Philadelphia: John Benjamins.
Dirven, René and John Taylor
1988 The conceptualisation of vertical space in English: The case of
“tall”. In Topics in Cognitive Linguistics, Brygida Rudzka-Ostyn
(ed.), 379-402. Amsterdam / Philadelphia: John Benjamins.
Dirven, René, Louis Goossens, Yvan Putseys and Emma Vorlat
1982 The Scene of Linguistic Action and its Perspectivization by SPEAK,
TALK, SAY, and TELL. Amsterdam / Philadelphia: John Benjamins.
30 Dylan Glynn

Diessel, Holger
2008 Iconicity of sequence. A corpus-based analysis of the positioning of
temporal adverbial clauses in English. Cognitive Linguistics 19:
457–482.
Divjak, Dagmar
2006 Ways of intending: A corpus-based cognitive linguistic approach to
near-synonyms in Russian. In Corpora in Cognitive Linguistics.
Corpus-based approaches to syntax and lexis, Stefan Th. Gries and
Anatol Stefanowitsch (eds.), 19–56. Berlin / New York: Mouton de
Gruyter.
2010 Structuring the Lexicon: A Clustered Model for Near-synonymy.
Berlin / New York: Mouton de Gruyter.
Divjak, Dagmar and Nick Fieller
in press Clustering linguistic data. In Polysemy and Synonymy. Corpus
methods and applications in Cognitive Linguistics, Dylan Glynn and
Justyna Robinson (eds.). Amsterdam / Philadelphia: John Benja-
mins.
Divjak, Dagmar and Stefan Th. Gries
2006 Ways of trying in Russian: Clustering behavioral profiles. Corpus
Linguistics and Linguistic Theory 2: 23–60.
2009 Corpus-based Cognitive Semantics: A contrastive study of phasal
verbs in English and Russian. In Studies in Cognitive Corpus Lin-
guistics, Barbara Lewandowska-Tomaszczyk and Katarzyna Dzi-
wirek (eds.), 273-296. Frankfurt/Main: Peter Lang.
Durkin, Kevin and Jocelyn Manning
1989 Polysemy and the subjective lexicon: Semantic relatedness and the
salience of intraword senses. Journal of Psycholinguistic Research
18: 577-612.
Dziwirek, Katarzyna and Barbara Lewandowska-Tomaszczyk
2009 Love and hate. Unique transitive emotions in English and Polish. In
Studies in Cognitive Corpus Linguistics, Barbara Lewandowska-
Tomaszczyk and Katarzyna Dziwirek (eds.), 297-318. Frank-
furt/Main: Peter Lang.
Fabiszak, Małgorzata, Anna Hebda, Iwona Kokorniak and Karolina Krawczak
in press The prefix semantics and grammar in the Polish verb ‘to think’. In
Polysemy and Synonymy. Corpus methods and applications in Cog-
nitive Linguistics, Dylan Glynn and Justyna Robinson (eds.).
Amsterdam / Philadelphia: John Benjamins.
Fillmore, Charles
1985 Frames and the Semantics of Understanding. Quaderni di Semantica
6: 222-254.
 Corpus-driven Cognitive Semantics 31

Fillmore, Charles and Beryl Atkins

1992 Toward a frame-based lexicon: The semantics of RISK and its
neighbours. In Frames, Fields, and Contrasts: New essays in seman-
tic and lexical organization, Adrienne Lehrer and Eva Kittay (eds.),
75-102. London: Lawrence Erlbaum Associates.
Fischer, Kerstin
2000 From Cognitive Semantics to Lexical Pragmatics: The functional
polysemy of discourse particles. Berlin / New York: Mouton de
Gruyter.
Geeraerts, Dirk
1993a Vagueness’s puzzles, polysemy’s vagaries. Cognitive Linguistics 4:
223-272.
1993b Generalised onomasiological salience. In Perspectives on Language
and Conceptualization (Special ed. Belgian Journal of Linguistics
8), Jan Nuyts and Eric Pederson (eds.), 43-56. Brussels: Presses
Universitaires de Bruxelles.
1998 The semantic structure of the indirect object in Dutch. In The Dative.
Vol. 2. Theoretical and contrastive studies, Willy van Langendonck
and William van Belle (eds.), 185-210. Amsterdam / Philadelphia:
John Benjamins.
2000 Salience phenomena in the lexicon. A typology. In Meaning and
Cognition, Liliana Albertazzi (ed.), 125-136. Amsterdam / Philadel-
phia: John Benjamins.
2005 Lectal data and empirical variation in Cognitive Linguistics. In Cog-
nitive Linguistics. Internal dynamics and interdisciplinary interac-
tions, Francisco José Ruiz de Mendoza Ibáñez and Sandra Peña
Cervel (eds.), 163-189. Berlin: Mouton de Gruyter.
2006a Words and Other Wonders: Papers on lexical and semantic topics.
Berlin / New York: Mouton de Gruyter.
2006b Methodology in Cognitive Linguistics. In Cognitive Linguistics:
Current applications and future perspectives, Gitte Kristiansen,
Michel Achard, René Dirven and Francisco J. Ruiz de Mendoza
Ibañez (eds.), 21-50. Berlin / New York: Mouton de Gruyter.
2007 Interview: Cognitive sociolinguistics and the sociology of Cognitive
Linguistics, by Juana Marín-Arrese, Annual Review of Cognitive
Linguistics 5: 289–305.
2010 Recontextualizing grammar: Underlying trends in thirty years of
Cognitive Linguistics. In Cognitive Linguistics in Action: From
theory to application and back, Elżbieta Tabakowska, Micha Choin-
ski and Łukasz Wiraszka (eds.), 71–102. Berlin / New York: Mou-
ton de Gruyter.
32 Dylan Glynn

Geeraerts, Dirk and Speelman, Dirk

2010 Heterodox concept features and onomasiological heterogeneity in
dialects. In Advances in Cognitive Sociolinguistics, Dirk Geeraerts,
Gitte Kristiansen and Yves Peirsman (eds.), 21-40. Berlin / New
York: Mouton de Gruyter.
Geeraerts, Dirk, Gitte Kristiansen and Yves Peirsman (eds.)
2010 Advances in Cognitive Sociolinguistics. Berlin / New York: Mouton
de Gruyter.
Geeraerts, Dirk, Stefan Grondelaers and Dirk Speelman
1999 Convergentie en Divergentie in de Nederlandse Woordenschat.
Amsterdam: Meertens Instituut.
Geeraerts, Dirk, Stefan Grondelaers and Peter Bakema
1994 The Structure of Lexical Variation. Meaning, naming, and context.
Berlin / New York: Mouton de Gruyter.
Gevaert, Caroline
2005 The anger is heat question: Detecting cultural influence on the con-
ceptualization of anger through diachronic corpus analysis. In Per-
spectives on Variation. Sociolinguistic, historical, comparative,
Nicole Delbecque, Johan van der Auwera and Dirk Geeraerts (eds.),
195-208. Berlin / New York: Mouton de Gruyter.
Gibbs, Raymond
1990 Psycholinguistic studies on the conceptual basis of idiomaticity.
Cognitive Linguistics 1: 417-451.
Gilquin, Gaëtanelle
2006 The place of prototypicality in corpus linguistics. Causation in the
hot seat. In Corpora in Cognitive Linguistics. Corpus-based ap-
proaches to syntax and lexis, Stefan Th. Gries and Anatol Ste-
fanowitsch (eds.), 159-191. Berlin / New York: Mouton de Gruyter.
2010 Corpus, Cognition and Causative Constructions. Amsterdam /
Philadelphia: John Benjamins.
Gilquin, Gaëtanelle (ed.)
2009 Corpora and Experimental Methods (Special ed. Corpus Linguistics
and Linguistic Theory 5). Berlin / New York: Mouton de Gruyter.
Gilquin, Gaëtanelle and Stefan Th. Gries
2009 Corpora and experimental methods: A state-of-the-art review. Cor-
pus Linguistics and Linguistic Theory 5: 1-26.
Givón, Talmy
1991 Markedness in grammar: distributional, communicative and cogni-
tive correlates of syntactic structure. Studies in Language 15: 335-
370.
 Corpus-driven Cognitive Semantics 33

Glynn, Dylan
2009 Polysemy, syntax, and variation. A usage-based method for Cogni-
tive Semantics. In New Directions in Cognitive Linguistics, Vyvyan
Evans and Stéphanie Pourcel (eds.), 77-106. Amsterdam / Philadel-
phia: John Benjamins.
2010 Synonymy, lexical fields, and grammatical constructions. A study in
usage-based Cognitive Semantics. In Cognitive Foundations of Lin-
guistic Usage-Patterns, Hans-Jörg Schmid and Susanne Handl
(eds.), 89-118. Berlin / New York: Mouton de Gruyter.
in press a Lexical synonymy and onomasiological structure in a usage-based
model of language. A multifactorial case study. In Polysemy and
Synonymy. Corpus methods and applications in Cognitive Linguist-
ics, Dylan Glynn and Justyna Robinson (eds.). Amsterdam /
Philadelphia: John Benjamins.
in press b Multiple correspondence analysis. Exploring correlations in multi-
factorial data. In Polysemy and Synonymy. Corpus methods and ap-
plications in Cognitive Linguistics, Dylan Glynn and Justyna Robin-
son (eds.). Amsterdam / Philadelphia: John Benjamins.
submitted Semantics of sociolinguistic variation. A quantitative study of dia-
lect effects on the polysemy of annoy. Annual Review of Cognitive
Linguistics.
forthc. Usage-based cognitive models. Lexical synonymy, conceptual struc-
ture and American LIBERTY. Cognitive Linguistics.
Glynn, Dylan and Justyna Robinson (eds.)
in press Polysemy and Synonymy. Corpus methods and applications in Cog-
nitive Linguistics. Amsterdam / Philadelphia: John Benjamins.
Glynn, Dylan and Karolina Krawczak
submitted Social cognition, Cognitive Grammar and corpora. A multifactorial
approach to subjectification. Language and Cognition
Goldberg, Adele
1995 Constructions. A Construction Grammar approach to argument
structure. London: University of Chicago Press.
Gonzalez-Marquez, Monica, Irene Mittleberg, Seanna Coulson and Michael
Spivey (eds.)
2007 Methods in Cognitive Linguistics. Amsterdam / Philadelphia: John
Benjamins.
Goossens, Louis
1990 Metaphtonymy: The interaction of metaphor and metonymy in figu-
rative expressions for linguistic action. Cognitive Linguistics 1: 323-
34.
34 Dylan Glynn

Gries, Stefan Th.

1999 Particle movement: A cognitive and functional approach. Cognitive
Linguistics 10: 105-145.
2001 A multifactorial analysis of syntactic variation: Particle movement
revisited. Journal of Quantitative Linguistics 8: 33-50.
2003 Multifactorial Analysis in Corpus Linguistics: A study of particle
placement. London: Continuum Press.
2006a Corpus-based methods and Cognitive Semantics: The many senses
of to run. In Corpora in Cognitive Linguistics. Corpus-based ap-
proaches to syntax and lexis, Stefan Th. Gries and Anatol Ste-
fanowitsch (eds.), 57–99. Berlin / New York: Mouton de Gruyter.
2006b Exploring variability within and between corpora: Some method-
ological considerations. Corpora 1: 109–51.
2008 Dispersions and adjusted frequencies in corpora. International Jour-
nal of Corpus Linguistics 13: 403-437.
Gries, Stefan Th. and Anatol Stefanowitsch (eds.)
2006 Corpora in Cognitive Linguistics. Corpus-based approaches to
syntax and lexis. Berlin / New York: Mouton de Gruyter.
Gries, Stefan Th. and Anatol Stefanowitsch
2004 Extending collostructional analysis: a corpus-based perspective on
‘alternations’. International Journal of Corpus Linguistics 9: 97-
129.
2010 Cluster analysis and the identification of collexeme classes. In Em-
pirical and Experimental Methods in Cognitive/Functional Re-
search, Sally Rice and John Newman (eds.), 73-90. Stanford, CA:
CSLI.
Gries, Stefan Th. and Dagmar Divjak (eds.)
forthc. Frequency Effects in Language. Berlin / New York: Mouton de
Gruyter.
Gries, Stefan Th. and Martin Hilpert
2008 The identification of stages in diachronic data: Variability-based
neighbor clustering. Corpora 3: 59-81.
Gries, Stefan Th., Beate Hampe and Doris Schönefeld
2005 Converging evidence: Bringing together experimental and corpus
data on the association of verbs and constructions. Cognitive Lin-
guistics 16: 635-676.
2010 Converging evidence II: More on the association of verbs and con-
structions. In Empirical and Experimental Methods in Cogni-
tive/Functional Research, Sally Rice and John Newman (eds.), 59-
72. Stanford: CSLI.
 Corpus-driven Cognitive Semantics 35

Gries, Stefan Th., Stefanie Wulff and Mark Davies (eds.)

2009 Corpus Linguistic Applications: Current studies, new directions.
Amsterdam: Rodopi.
Grondelaers, Stefan, Dirk Geeraerts and Dirk Speelman
2007 A case for a cognitive corpus Linguistics. In Methods in Cognitive
Linguistics, Monica Gonzalez-Marquez, Irene Mittleberg, Seanna
Coulson and Michael Spivey (eds.), 149-169. Amsterdam / Philadel-
phia: John Benjamins.
Grondelaers, Stefan, Dirk Speelman and Dirk Geeraerts
2008 National variation in the use of er “there”. Regional and diachronic
constraints on cognitive explanations. In Cognitive Sociolinguistics:
Language variation, cultural models, social systems, Gitte Kristian-
sen and René Dirven (eds.), 153-204. Berlin / New York: Mouton de
Gruyter.
Hanegreefs, Hilde
2004 The Dutch change-of-state copula ‘worden’ and its Spanish counter-
parts: A matter of aspect and voice. Annual Review of Cognitive
Linguistics 1: 29.
Heylen, Kris
2005 A quantitative corpus study of German word order variation. In
Linguistic Evidence: Empirical, theoretical and computational per-
spectives, Stephan Kepser and Marga Reis (eds.), 241-264. Berlin /
New York: Mouton de Gruyter.
Heylen, Kris, José Tummers and Dirk Geeraerts
2008 Methodological issues in corpus-based Cognitive Linguistics. In
Cognitive Sociolinguistics: Language variation, cultural models,
social systems, Gitte Kristiansen and René Dirven (eds.), 91-128.
Berlin / New York: Mouton de Gruyter.
Hilpert, Martin
2008a Germanic Future Constructions. A usage-based approach to lan-
guage change. Amsterdam / Philadelphia: John Benjamins.
2008b The English comparative – language structure and language use.
English Language and Linguistics 12: 395–417.
2009 The German mit-predicative construction. Constructions and
Frames 1: 29-55.
in press Collostructional Analysis: Measuring associations between construc-
tions and lexical elements. In Polysemy and Synonymy. Corpus
methods and applications in Cognitive Linguistics, Dylan Glynn and
Justyna Robinson (eds.). Amsterdam / Philadelphia: John Benja-
mins.
36 Dylan Glynn

Janda, Laura
1993 A Geography of Case Semantics: The Czech dative and the Russian
instrumental. Berlin / New York: Mouton de Gruyter.
Janda, Laura and Valery Solovyev
2009 What constructional profiles reveal about synonymy: A case study
of the Russian words for SADNESS and HAPPINESS. Cognitive Lin-
guistics 20: 367-393.
in press Antonyms as types of synonyms: Russian emotion terms. In Polys-
emy and Synonymy. Corpus methods and applications in Cognitive
Linguistics, Dylan Glynn and Justyna Robinson (eds.). Amsterdam /
Philadelphia: John Benjamins.
Jing-Schmidt, Zhuo and Stefan Th. Gries
2009 Schematic meaning and pragmatic inference: The Mandarin adverbs
hai, you, and zai. Corpora 4: 33-70.
Jones, Steven, Carita Paradis, Lynne Murphy and Caroline Willners
2007 Googling for ‘opposites’: A web-based study of antonym canonicity.
Corpora 2: 129-155.
Kabata, Kaori and Jeong-Hwa Lee
2010 ‘Come’ and ‘go’ serial verbs in Japanese and Korean: A corpus-
based comparative study. In Empirical and Experimental Methods in
Cognitive/Functional Research, Sally Rice and John Newman (eds.),
109 -126. Stanford: CSLI.
Klavan, Jane
in press A multifactorial analysis of grammatical synonymy: The case of the
Estonian adessive and the adposition peal ‘on’. In Polysemy and
Synonymy. Corpus methods and applications in Cognitive Linguist-
ics, Dylan Glynn and Justyna Robinson (eds.). Amsterdam /
Philadelphia: John Benjamins.
Kövecses, Zoltán
1986 Metaphors of Anger, Pride, and Love. A lexical approach to the
structure of concepts. Amsterdam / Philadelphia: John Benjamins.
Kristiansen, Gitte and René Dirven (eds.)
2008 Cognitive Sociolinguistics: Language variation, cultural models,
social systems. Berlin / New York: Mouton de Gruyter.
Lakoff, George
1982 Categories. An essay in Cognitive Linguistics. In Linguistics in the
Morning Calm, Linguistic Society of Korea (ed.), 139-194. Seoul:
Hanshin.
1987 Women, Fire, and Dangerous Things. What categories reveal about
the mind. London: University of Chicago Press.
 Corpus-driven Cognitive Semantics 37

Langacker, Ronald
1987 Foundations of Cognitive Grammar. Vol. 1. Theoretical prerequi-
sites. Stanford: Stanford University Press.
1988 A usage-based model. In Topics in Cognitive Linguistics, Brygida
Rudzka-Ostyn (ed.), 127-161. Amsterdam / Philadelphia: John Ben-
jamins.
2000 Grammar and Conceptualization. Berlin / New York: Mouton de
Gruyter.
Lehrer, Adrienne
1982/2000 Wine and Conversation. Bloomington: Indiana University Press/
Oxford: Oxford University Press.
Levshina, Natalia, Dirk Speelman and Dirk Geeraerts
in press Dutch causative constructions: Quantification of meaning and mean-
ing of quantification. In Polysemy and Synonymy. Corpus methods
and applications in Cognitive Linguistics, Dylan Glynn and Justyna
Robinson (eds.). Amsterdam / Philadelphia: John Benjamins.
Lewandowska-Tomaszczyk, Barbara and Katarzyna Dziwirek (eds.)
2009 Studies in Cognitive Corpus Linguistics. Frankfurt/Main: Lang.
Majid, Asifa and Melissa Bowerman
2007 Cutting and breaking events: A cross-linguistic perspective. Cogni-
tive Linguistics 18: 133-152.
Marzo, Stefania, Kris Heylen and Gert De Sutter (eds.)
2010 Corpus Studies in Contrastive Linguistics (Special ed. International
Journal of Corpus Linguistics 15). Amsterdam / Philadelphia: John
Benjamins.
Mukherjee, Joybrato
2005 English Ditransitive Verbs: Aspects of theory, description and a
usage-based model. Amsterdam / New York: Rodopi.
Newman, John
2008a Spoken corpora: Rationale and application. Taiwan Journal of Lin-
guistics 6: 27-58.
2008b Aiming low in linguistics: Low-level generalizations in corpus-
based research. 11th International Symposium on Chinese Lan-
guages and Linguistics, Chiao Tung University, Taiwan.
2009 English posture verbs: An experientially grounded approach. Annual
Review of Cognitive Linguistics 7: 30–57.
Newman, John and Sally Rice
2004a Aspect in the making: A corpus analysis of English aspect-marking
prepositions. In Language, Culture and Mind, Suzanne Kemmer and
Michel Achard (eds.), 313-327. Stanford: CSLI.
38 Dylan Glynn

2004b Patterns of usage for English SIT, STAND, and LIE: A cognitively-
inspired exploration in corpus linguistics. Cognitive Linguistics 15:
351-396.
2006 Transitivity schemas of English EAT and DRINK in the BNC. In Cor-
pora in Cognitive Linguistics. Corpus-based approaches to syntax
and lexis, Stefan Th. Gries and Anatol Stefanowitsch (eds.), 225-
260. Berlin / New York: Mouton de Gruyter.
Noël, Dirk and Timothy Colleman
2010 Believe-type raising-to-object and raising-to-subject verbs in English
and Dutch: A contrastive investigation in diachronic Construction
Grammar. Corpus Studies in Contrastive Linguistics, Ste-
fania Marzo, Kris Heylen and Gert De Sutter (eds.), 157–182.
Amsterdam / Philadelphia: John Benjamins.
Nokkonen, Soili
2006 The semantic variation of NEED TO in four recent British English
corpora. International Journal of Corpus Linguistics 11: 29-71.
Paradis, Carita
2003 Is the notion of linguistic competence relevant in Cognitive Linguist-
ics? Annual review of Cognitive Linguistics 1: 247-271.
Peirsman, Yves, Kris Heylen and Dirk Geeraerts
2010 Applying word space models to sociolinguistics. In Advances in
Cognitive Sociolinguistics, Dirk Geeraerts, Gitte Kristiansen and
Yves Peirsman (eds.), 111-138. Berlin / New York: Mouton de
Gruyter.
Rice, Sally and John Newman (eds.)
2010 Empirical and Experimental Methods in Cognitive/Functional Re-
search. Stanford: CSLI.
Rudzka-Ostyn, Brygida
1989 Prototypes, schemas, and cross-category correspondences: The case
of ask. In Prospects and Problems of Prototype Theory (Special ed.
Linguistics 27), Dirk Geeraerts (ed.), 613-661. Berlin / New York:
Mouton de Gruyter.
1995 Metaphor, schema, invariance: The case of verbs of answering. In
By Word of Mouth. Metaphor, Metonymy, and Linguistic Action
from a Cognitive Perspective, Louis Goossens, Paul Pauwels,
Brygida Rudzka-Ostyn, Anne-Marie Simon-Vandenbergen and
Johan Vanparys (eds.), 205-244. Amsterdam / Philadelphia: John
Benjamins.
1996 The Polish Dative. In The Dative. Vol. 1. Descriptive studies. Wil-
liam van Belle and Willy van Langendonck (eds.), 341-394.
Amsterdam / Philadelphia: John Benjamins.
 Corpus-driven Cognitive Semantics 39

Sandra, Dominiek and Sally Rice

1995 Network analysis of prepositional meaning: mirroring whose mind -
the linguist’s or the language user’s? Cognitive Linguistics 6: 89-
130.
Schmid, Hans-Jörg
1993 Cottage and Co., Idea, Start vs. Begin. Die Kategorisierung als
Grundprinzip einer differenzierten Bedeutungsbeschreibung. Tübin-
gen: Niemeyer.
2000 English Abstract Nouns as Conceptual Shells. From corpus to cog-
nition. Berlin / New York: Mouton de Gruyter.
Schulze, Rainer
1988 A short story of down. In Understanding the Lexicon. Meaning,
sense, and world knowledge in lexical semantics, Werner Hüllen and
Rainer Schulze (eds.), 395-414. Tübingen: Niemeyer.
Soares da Silva, Augusto
2010 Measuring and parameterizing lexical convergence and divergence
between European and Brazilian Portuguese. In Advances in Cogni-
tive Sociolinguistics, Dirk Geeraerts, Gitte Kristiansen and Yves
Peirsman (eds.), 41-84. Berlin / New York: Mouton de Gruyter.
Speelman, Dirk
in press Modeling near-synonymy with logistic regression. In Polysemy and
Synonymy. Corpus methods and applications in Cognitive Linguist-
ics, Dylan Glynn and Justyna Robinson (eds.). Amsterdam /
Philadelphia: John Benjamins.
Speelman, Dirk and Dirk Geeraerts
2010 Causes for causatives: the case of Dutch ‘doen’ and ‘laten’. In Cau-
sal Categories in Discourse and Cognition, Ted Sanders and Eve
Sweetser (eds.), 173-204. Berlin / New York: Mouton de Gruyter.
Speelman, Dirk, José Tummers and Dirk Geeraerts
2009 Lexical patterning in a Construction Grammar: The effect of lexical
co-occurrence patterns on the inflectional variation in Dutch attribu-
tive adjectives. Constructions and Frames 1: 87–118.
Speelman, Dirk, Stefan Grondelaers and Dirk Geeraerts
2003 Profile-based linguistic uniformity as a generic method for compar-
ing language varieties. Computers and the Humanities 37: 317-337.
Stefanowitsch, Anatol
2006 Negative evidence and the raw frequency fallacy. Corpus Linguistics
and Linguistic Theory 2: 61-77.
2008 Negative entrenchment: A usage-based approach to negative evi-
dence. Cognitive Linguistics 19: 513-531.
40 Dylan Glynn

Stefanowitsch, Anatol and Stefan Th. Gries

2003 Collostructions: Investigating the interaction of words and construc-
tions. International Journal of Corpus Linguistics 8: 209-243.
2008 Register and constructional semantics: A collostructional case study.
In Cognitive Sociolinguistics, Gitte Kristiansen and René Dirven
(eds.), 129–152. Berlin / New York: Mouton de Gruyter.
Stefanowitsch, Anatol and Stefan Th. Gries (eds.)
2006 Corpus-Based Approaches to Metaphor and Metonymy. Berlin /
New York: Mouton de Gruyter.
Szelid, Veronika and Dirk Geeraerts
2008 Usage-based dialectology. Emotion concepts in the Southern Csango
dialect. Annual Review of Cognitive Linguistics 6: 23-49.
Szmrecsanyi, Benedikt
2003 Be going to versus will/shall: Does syntax matter? Journal of Eng-
lish Linguistics 31: 295-323.
2006 Morphosyntactic Persistence in Spoken English. A corpus study at
the intersection of variationist sociolinguistics, psycholinguistics,
and discourse analysis. Berlin / New York: Mouton de Gruyter.
2010 The English genitive alternation in a Cognitive Sociolinguistics
perspective. In Advances in Cognitive Sociolinguistics, Dirk Geer-
aerts, Gitte Kristiansen and Yves Peirsman (eds.), 141-166. Berlin /
New York: Mouton de Gruyter.
Talmy, Leonard
1985 Force dynamics in language and cognition. Cognitive Science 12:
49-100.
2000 Toward a Cognitive Semantics. Vol. 1. Concept structuring systems.
Cambridge, MA: MIT Press.
2008 Foreword. In Methods in Cognitive Linguistics, Monica Gonzalez-
Marquez, Irene Mittleberg, Seanna Coulson and Michael Spivey
(eds.), xxi – xxi. Amsterdam / Philadelphia: John Benjamins.
Tomasello, Michael
2000 First steps toward a usage-based theory of language acquisition.
Cognitive Linguistics 11: 61–82.
Tummers, José, Dirk Speelman and Dirk Geeraerts
2005 Inflectional variation in Belgian and Netherlandic Dutch: A usage-
based account of the adjectival inflection. In Perspectives on Vari-
ation. Sociolinguistic, historical, comparative, Nicole Delbecque,
Johan van der Auwera and Dirk Geeraerts (eds.), 93-110. Berlin /
New York: Mouton de Gruyter.
Tummers, Jose, Kris Heylen and Dirk Geeraerts
2005 Usage-based approaches in Cognitive Linguistics: A technical state
of the art. Corpus Linguistics and Linguistic Theory 1(2): 225-261.
 Corpus-driven Cognitive Semantics 41

Wiechmann, Daniel
2008a On the computation of collostruction strength: Testing measures of
association as expressions of lexical bias. Corpus Linguistics and
Linguistic Theory 4: 253-290.
2008b Sense-contingent lexical preferences and early parsing decisions.
Cognitive Linguistics 19: 439-455.
Wierzbicka, Anna
1985 Lexicography and Conceptual Analysis. Ann Arbor: Karoma.
1990 Prototypes ‘save’: on the uses and abuses of the notion of ‘proto-
type’ in linguistics and related fields. In Meanings and Prototypes.
Studies on linguistic categorization, Savas Tsohatzidis (ed.), 347-
367. London: Routledge.
Wulff, Stefanie
2003 A multifactorial corpus analysis of adjective order in English. Inter-
national Journal of Corpus Linguistics 8: 245-82.
2006 Go-V vs. go-and-V in English: A case of constructional synonymy?
In Corpora in Cognitive Linguistics. Corpus-based approaches to
syntax and lexis, Stefan Th. Gries and Anatol Stefanowitsch (eds.),
101-126. Berlin / New York: Mouton de Gruyter.
2009a Rethinking Idiomaticity. A usage-based approach. London: Con-
tinuum.
2009b Converging evidence from corpus and experimental data to capture
idiomaticity. Corpus Linguistics and Linguistics Theory 5: 131-159.
Wulff, Stefanie, Anatol Stefanowitsch and Stefan Th. Gries
2007 Brutal Brits and persuasive Americans: Variety-specific meaning
construction in the into-causative. In Aspects of Meaning Construc-
tion, Günter Radden, Klaus-Michael Köpcke, Thomas Berg and
Peter Siemund (eds.), 265-281. Amsterdam / Philadelphia: John
Benjamins.
Zeschel, Arne (ed.)
2008 Usage-Based Approaches to Language Processing and Representa-
tion (Special ed. Cognitive Linguistics 19). Berlin / New York: Mou-
ton de Gruyter.
Zeschel, Arne
2009 What’s (in) a construction? In New directions in Cognitive Linguist-
ics, Vyvyan Evans and Stéphanie Pourcel (eds.), 185 - 200. Amster-
dam / Philadelphia: John Benjamins.
Quantitative Methods in Cognitive Semantics
Introduction to the volume

Kerstin Fischer

1. The Empirical Turn

Cognitive linguistics currently experiences a trend towards the use of em-

pirical methods; it can be observed that conferences fill up increasingly
with empirical studies supporting and extending the conceptual work in
Cognitive Linguistics. At the same time, several volumes with the explicit
aim to lay the foundations for empirical investigations have recently ap-
peared (e.g. Gries and Stefanowitsch 2006; Kristiansen, Achard, Dirven
and Ruiz de Mendoza Ibáñez 2006; González-Márquez, Mittelberg, Coul-
son and Spivey 2007). These new developments concern the relationship of
cognitive linguistic findings to other empirical disciplines, such as psychol-
ogy, corpus- and neurolinguistics, on the one hand and the attempt to
achieve additional methodological rigour on the other. With respect to the
former, Gibbs (2007), for instance, suggests Cognitive Linguistics to in-
corporate more psycholinguistic findings in order to live up to the require-
ment of cognitive reality. Stefanowitsch (2009) holds that in order for Cog-
nitive Linguistics to become a cognitive science, it needs to pay more
attention to its interfaces with other disciplines, which includes attention to
methodology and to generally accepted criteria of scientific investigation,
such as intersubjectivity, representativity, reliability, and validity. With
regard to the latter, Tummers et al. (2005) suggest that more methodologi-
cal rigour be needed in Cognitive Linguistics, and Geeraerts (2006), for
instance, explicitly argues for an ‘empirical turn’ in Cognitive Linguistics.
He evaluates the current situation as a ‘theoretical chaos’, albeit a creative
one, which needs to give room to a situation in which competing ap-
proaches can be objectively compared (Geeraerts 2006: 21). The use of
quantitative methodologies in Cognitive Semantics presented in this vol-
1
ume is similarly an attempt at studying Cognitive Semantics empirically, at
connecting Cognitive Semantics with research in other disciplines, espe-
cially corpus linguistics and usage-based approaches to language, and at
introducing more methodological rigour into the discipline.
44 Kerstin Fischer

The question now is: How can meaning in Cognitive Semantics be in-
vestigated by means of quantitative methods? Or, as Geeraerts puts it in this
volume: “how can meaning, the most qualitative of all linguistic features,
be expressed in numbers?”
In fact there are many who consider the use of quantitative methodolo-
gies to be too problematic to produce valid results. Already the definition of
the numerator, the item under consideration, may not be trivial; for in-
stance, it may be polysemous, its interpretation may be context-dependent
(cf. Bednarek 2008), and even different word forms of the item in question
may have different uses (cf. Newman, this vol.). The definition of the de-
nominator, the actual and potential contexts in which the item under con-
sideration may occur (cf. Schmid, this vol.), and the domain, the context in
which the data were observed, i.e. the corpus under consideration, may be
similarly problematic (Fischer, this vol.). Especially in a constructivist ap-
proach such as Cognitive Linguistics, categories cannot be assumed to be
objectively given. Heritage (1995: 400-402) illustrates the problem for the
occurrence of the interjection ‘oh’ as the third turn after responses to ques-
tions. While “the third turn after replies to questions” may appear as a use-
ful operationalisation, in practice it is often not trivial to identify objec-
tively which utterances constitute questions and which responses.
Moreover, the problems increase in the realm of meaning; given the
cognitive nature of meanings in Cognitive Semantics, the use of quantita-
tive methodology is not obvious. In general, quantitative methods introduce
an external, objectifying perspective on the object of study (Heritage 1995:
406). While this ensures intersubjectivity in the scientific endeavour, it may
be problematic for the study of meaning as “the most subjective of all lin-
guistic phenomena” (Geeraerts, this vol.).
The current volume illustrates some ways in which applying corpus lin-
guistic methods and statistical analysis to the study of meaning may be
fruitful; it presents practical solutions for the methodological problems
outlined above and demonstrates procedures and methods for the empirical
study of meaning. However, it also addresses the problems arising in a
quantitative approach to meaning and the implications the use of quantita-
tive methodologies may have for the development of the discipline. Thus,
the volume reflects the struggle of a relatively young discipline determining
its future course of action.
 Quantitative Methods in Cognitive Semantics 45

2. Quantitative methods in Cognitive Semantics

In Cognitive Semantics, meaning is understood as a cognitive phenomenon;

Lakoff (1987), for instance, outlines the cognitive linguistic program in
contrast to so-called objectivist approaches to language (see also Langacker
1987: 5; Fillmore 1975). Thus, meaning in Cognitive Linguistics is taken to
correspond to dynamic, context-sensitive cognitive construal, for instance:
The term conceptualization is interpreted broadly as embracing any kind of
mental experience. It subsumes (a) both established and novel conceptions;
(b) not only abstract or intellectual ‘concepts’ but also sensory, motor, and
emotive experience; (c) conceptions that are not instantaneous but change or
unfold through processing time; and (d) full apprehension of the physical,
social, cultural, and linguistic context. Thus, far from being either static or
solipsistic, conceptualization is viewed as the dynamic activity of embodied
minds interacting with their environments. (Langacker 1998: 3)
Crucial aspects of such construal are, according to Langacker (1998),
granularity, metaphor, perspective, and prominence, in particular profiling.
These processes guide the way in which meaning is cognitively construed
in different linguistic structures.
Langacker's definition of meaning in Cognitive Semantics also means
that linguistic meaning and world knowledge cannot be separated categori-
cally. This view is generally shared in Cognitive Linguistics; for instance,
in various seminal papers (e.g. 1975, 1982, 1985; Fillmore and Atkins
1992) Fillmore outlines the cognitive semantic approach to meaning that he
calls a semantics of understanding. This comprises “what one knows by
virtue of being a speaker of the language” (Fillmore 1985: 252). Fillmore
argues that a semantic theory should aim to account for “the relationship
between linguistic texts, the context in which they are instanced, and the
process and products of their interpretation” (1985: 222). That is, “in de-
termining the situated meanings of uses of the sentence, one interprets the
sentence’s conventional meaning with its linguistic and extralinguistic con-
text” (1985: 233).
Cognitive semantics thus focuses on conceptualization and understand-
ing; yet, it also makes use of the notion of reference, although this notion is
heavily relied on in formal semantic approaches, which are generally re-
jected from a cognitive linguistic perspective. Reference may provide the
central anchor point to compare, for instance, different construals (Berlin
and Kay 1969, Geeraerts, Grondelaers and Bakema 1994; Langacker 2008:
ch.9).
46 Kerstin Fischer

Finally, usage plays an important role in Cognitive Semantics. Espe-

cially Langacker has been advocating the usage-based approach (e.g. 1987,
1988, 2000, 2008) that holds that meaning emerges from use. Given
Fillmore’s above definition of the semantics of understanding, expressions
are taken to receive their meanings in contexts of use. Such a perspective
entails different degrees of schematization, the study of creativity and ana-
logical reasoning, and opens up for the possibility that even different word
forms may be entrenched differently. In the same way as meaning emerges
from use, language use is taken to be determined by the conceptual mean-
ing by an item, and, in the same way, its reference is determined by its con-
ceptual content.
To sum up, Cognitive Semantics involves four different aspects of
meaning (see Figure 1): conceptualisation, usage, world knowledge and
reference. These are taken to be highly interrelated and in many ways in-
separable, such that context, meaning, reference and use are taken to co-
determine each other.

conceptualisation reference

usage context /
world knowledge
Figure 1. Meaning in Cognitive Semantics
These different aspects of meaning lend themselves to study by means of
quantitative methodologies to different degrees. While conceptualisation is
generally inaccessible to direct scientific study, usage can be investigated
intersubjectively in collections of usage events, such as linguistic corpora.
Correspondingly, several papers in this volume approach Cognitive Se-
mantics with corpus linguistic methods; many researchers investigate lan-
guage usage therefore not in its own right but because they take it as an
indicators for something else (more interesting) (cf. Dirk Geeraerts, this
volume, and Anatol Stefanowitsch, this vol.), namely the conceptual mean-
ings of the items and structures under consideration. Since language use is
objectively accessible, it can serve as operationalisation of more subjective
factors if the link between the two factors can be unambiguously estab-
lished. Furthermore, in contrast to many experimental techniques, which
 Quantitative Methods in Cognitive Semantics 47

study conceptualisation in off-line experiments, corpora represent the ‘re-

sults’ of on-line language production (cf. Gries and Divjak 2009).
Accordingly, Stefan Gries and Dagmar Divjak argue in this volume that
the ‘behavioural profile’ of a word, i.e. a quantified description of its distri-
bution in large collections of data regarding a set of linguistic variables,
reflects its semantics. The underlying assumption is that the exact proper-
ties of a concept determine the word’s usage as it is manifest in corpora. In
their approach, the quantitative behavioural profile may then serve as the
basis for further quantitative investigation.
Unlike Stefan Gries and Dagmar Divjak, other researchers do not oper-
ate on the quantified results of the quantitative methods directly but rather
investigate those linguistic items further that the quantitative analysis iden-
tified as significantly attracted to the linguistic structure under consider-
ation (see Timothy Coleman, Martin Hilpert and Stefan Fuhs, this volume,
as well as Stefanowitsch and Gries to appear). Similarly, Kerstin Fischer
(this vol.) uses correlating features of the linguistic contexts in comparable
corpora as indicators for speakers’ conceptualisations of the communica-
tion situation, which in turn provide evidence for the functions of the con-
structions under consideration. A fourth example of quantitative methods
focusing on language use is Hans-Joerg Schmid’s investigation of the rela-
tionship between objectively observable frequency and the cognitive lin-
guistic concept of entrenchment. Entrenchment, like conceptionalization, is
not objectively accessible, and thus needs to be studied indirectly. Schmid’s
investigation concerns the question whether frequency, as an observable
feature of language use, can be interpreted as an indicator of the mental
phenomenon entrenchment.
However, to employ quantitative methods to investigate language use in
order to study meaning is only one of several uses of quantitative methods
in Cognitive Semantics illustrated in this volume. Dirk Geeraerts, for in-
stance, argues that quantifying, and even statistical, procedures are merely a
consequence of using empirical methods in general (this vol.).
A very different use of quantitative methodology is illustrated in Ste-
fanie Wulff’s paper; she uses quantitative methods to develop a mathemati-
cal model that describes and predicts compositionality in V-NP construc-
tions. That is, she uses results from corpus linguistic investigations to
quantify the contributions of the words inside a construction and its degree
of compositionality and then describes their relationship in a mathematical
model.
48 Kerstin Fischer

Dagmar Divjak’s paper illustrates another use of quantitative methods in

Cognitive Semantics; the author uses logistic regression to measure the
influence of particular factors in linguistic choice. That is, the quantitative
procedure she chooses allows her to identify those variables that predict a
particular linguistic form best, enabling her to distinguish between compet-
ing linguistic analyses of the phenomenon under consideration.
Finally, quantitative methods may also be helpful for investigating gen-
eral linguistic issues such as the determinants of meaning extension. Arne
Zeschel illustrates how quantitative methods can be applied to address the
relationship between established and novel semantic types in order to ex-
plain the productivity of certain word senses, i.e. the degree with which
they allow semantic extensions and analogical formations. By showing that
there is a high statistical correlation between the numbers of established
semantic types and novel instantiations, he provides empirical support for
an important theoretical concept.
To conclude, quantitative methodology may play several different roles
in Cognitive Semantics, supporting the scientific process in numerous dif-
ferent ways. While the most prominent use is the investigation of language
usage to identify conceptual meanings or entrenchment, quantitative meth-
ods may also serve to model findings and to contribute to theory develop-
ment by providing means to test hypotheses, to evaluate different theoreti-
cal models and to identify those factors that describe the observable
patterning of the data best.

3. Quantitative methods in Cognitive Semantics

Both John Newman and Hans-Joerg Schmid discuss (in this vol.) the prob-
lems of introducing methods into Cognitive Linguistics which originate
from another discipline and which may be too complicated for cognitive
linguists at large or which may not be psycholinguistically plausible. This
issue is similar to the role of psycholinguistic methods in Cognitive Lin-
guistics discussed by Gibbs (2007). Gibbs holds that even though Cognitive
Linguistics should incorporate psycholinguistic findings more, cognitive
linguists should not all try to become psychologists now; instead, the dif-
ferent competences should remain in the different disciplines. Applying this
proposal to quantitative Cognitive Semantics would mean that cognitive
linguists do what they have been doing for the last thirty years (which is
very much based on linguistic intuitions, cf. Geeraerts and Stefanowitsch in
 Quantitative Methods in Cognitive Semantics 49

this vol.) and leave it to a statistician to test cognitive linguistic hypotheses

in quantitative investigations.
None of the authors of this volume suggest this, and on the contrary,
many claim that Cognitive Semantics needs to rely on empirical methods
more, while others demonstrate the usefulness of such methods for various
problem areas in Cognitive Semantics. However, it is important to bear in
mind that the methodological discussion needs to be understandable to the
‘end user’, i.e. the cognitive linguist, as John Newman argues. The current
volume tries to bridge this gap by making a set of quantitative methods
more accessible to the reader and by illustrating their advantages and uses.
Moreover, most of the methods employed in this volume are rather ‘stan-
dard’ procedures such that they are available with most office software or
free to download; in any case, the introduction provided here is intended to
promote at least a passive understanding of the methods for the interested
reader.
The most commonly used method in this volume is the collostructional
method. This method covers in fact three different procedures, Collexeme
Analysis (Stefanowitsch and Gries 2003), Distinctive Collexeme Analysis
(Gries and Stefanowitsch 2004) and Covarying Collexeme Analysis (Ste-
fanowitsch and Gries 2005). The method used in the volume by Timothy
Coleman, Martin Hilpert and Stefan Fuhs is the collexeme analysis. In this
method, the frequency of a word in a given construction is compared to its
occurrence in other constructions and to the frequency of the construction
in the whole corpus. The resulting measure, the collostructional strength,
describes the association of a word in a construction and thus can serve as
starting point for further quantitative and qualitative investigations. This
method is also discussed in Hans-Joerg Schmid’s article in detail in com-
parison with some other measures of frequency in a corpus.
In addition to Collexeme Analysis, other statistical procedures are em-
ployed in the volume. For instance, Stefan Gries and Dagmar Divjak apply
a method from unsupervised learning to the quantitative distributions they
identify, cluster analysis. In this method, all instances in a data set are com-
pared to all others. The algorithm then determines a distance between each
data point and sorts the instances into a hierarchical set of classes, i.e. in a
tree-like representation, the dendrogram, where spatial proximity represents
similarity. This method can, for instance, be used to organise results from a
Collexeme Analysis into semantic classes (cf. Gries and Stefanowitsch
2010) or to detect prototypical word classes (Gries and Divjak, this vol.).
50 Kerstin Fischer

Another statistical method used in this volume is Correlation Analysis;

correlation describes the strength of the relationship between two variables,
suggesting possible causal relationships. The correlation coefficient used in
the papers by Arne Zeschel and Kerstin Fischer in this volume is the most
common one, the Pearson product moment coefficient, which describes the
strength of a linear relationship between two data sets. For instance, in a
strong positive correlation, if x is high, y will also be high; alternatively, in
a strong negative correlation, if x is high, y will be low. The Pearson pro-
duct moment correlation coefficient is part of most standard statistical an-
alysis software.
Moreover, Dagmar Divjak uses logistic regression in her study. Regres-
sion analysis measures the influence of a particular variable in a set of in-
dependent variables. That is, it analyses the contribution of a particular
variable to the behaviour of an independent variable if all other variables
are kept constant. It can thus be used as an exploratory method, identifying
which independent variable influences the results of the dependent variable
the most, but it is also a powerful technique from confirmatory statistics.
Dagmar Divjak uses this method to identify the contribution of a set of
hypothesized factors for the absence or presence of imperfective versus
perfective aspect in Russian.
The methods used and discussed in this volume do not constitute an ex-
haustive list of quantitative methods that may be usefully applied to Cogni-
tive Semantics; many other statistical (e.g. Glynn 2009 and this vol.) and
computational (e.g. Fischer 2000; Elman 2006) methods have been used in
the area of Cognitive Semantics, and so this volume may only illustrate a
commonly used subset of quantitative methods for Cognitive Semantics.
The focus of the volume is rather on determining the potential of quantita-
tive methodologies for the description of meaning and to determine the path
Cognitive Semantics is going to take in the near future.

4. Structure of the volume

The volume addresses the use of quantitative corpus methodologies for the
study of Cognitive Semantics. It intends to combine high quality, up-to-
date research on a broad range of cognitive semantic questions.
The volume starts with an outline of the history and principles of corpus
Cognitive Semantics. Following this overview, it presents a discussion of
those assumptions that may so far have prevented cognitive semanticists
 Quantitative Methods in Cognitive Semantics 51

from engaging in quantitative methodologies. Particularly for the domain of

semantics, the use of quantitative methodologies is not obvious, and espe-
cially the first two and the last two papers address the particular method-
ological problems arising. The first article sets the scene for rigorous hy-
pothesis formation and testing, which is put to use in the following articles.
The second article concerns some detailed methodological issues a cogni-
tive semantician using quantitative methodologies should bear in mind.
Following these stage-setting articles is a set of articles that address
central concepts in semantic research: semantic creativity, situation-
dependency, force-dynamics and coercion. These articles show how the
discussion of these central semantic concepts, which also lie at the core of
Cognitive Linguistics, can profit from a quantitative approach.
The following four articles provide in depth studies of selected method-
ological issues in Cognitive Semantics, combining detailed empirical inves-
tigation with methodological discussion. They concern the study of compo-
sitionality, the relationship between aspect and modality, lexical semantics
and the multifaceted semantics of an argument structure construction re-
spectively, which are discussed with respect to a variety of different lan-
guages.
Finally, the volume is concluded by two theoretical articles; the first one
addresses several reservations previously expressed against the use of quan-
titative methods in Cognitive Semantics and proposes the notion of behav-
ioural profile to frame theoretically how quantitative approaches to experi-
ential categories, such as meaning, may be investigated. The second one
outlines a general approach to empirical Cognitive Semantics.
In particular, the volume starts out with a stage-setting article by Dirk
Geeraerts, entitled “The Doctor and the Semantician” in which he first
addresses the question whether meaning can be studied empirically at all
and then discusses false conceptions why scholars may think they need not
address semantics using empirical methodologies. He continues by suggest-
ing to identify indices for the semantic concept in question that are easier to
establish than the concept under consideration itself. From there, he out-
lines a scientific procedure at which scholars in empirical Cognitive Se-
mantics may orient. He suggests to proceed in cycles of hypotheses and
testing, developing the theory in several steps.
John Newman’s “Balancing Acts: Empirical Pursuits in Cognitive Lin-
guistics” addresses some of the issues that scholars intending to use quanti-
tative methods need to consider. In particular, he considers those assump-
tions often not addressed in corpus-linguistic research. For instance, while
52 Kerstin Fischer

the usage-based approach is spreading in Cognitive Semantics, the concept

of ‘situated instances of language use’ is generally restricted to linguistic
cotext and excludes paralinguistic and suprasegmental linguistic features,
non-verbal behaviours and aspects of the context of situation. Moreover,
large-scale corpus-linguistic investigations usually concern written rather
than conversational language. John Newman continues by considering the
problems of ambiguous tags and of lemma-based corpus queries which
often neglect that different word forms may be entrenched differently. In
all, he argues for a balance between different kinds of methodologies.
Some of these considerations recur in Hans-Joerg Schmid’s “Does
Frequency in Text Instantiate Entrenchment in the Cognitive System?”; he
discusses different measures for frequency and compares calculations of
collostructional strength (Stefanowitsch and Gries 2003) with the possibly
more intuitive measures attraction and reliance; he concludes that subjec-
tive judgements may enter the calculations of the seemingly objective
measure frequency. Furthermore, he argues that the relationship between
frequency and entrenchment is more complex than usually assumed and
still in need of further research that integrates both corpus-linguistic and
psycholinguistic findings.
The theoretical section on corpus methods in Cognitive Semantics is fol-
lowed by a section in which the authors address specific theoretical issues.
Stefan Fuhs discusses in his article “Aspectual Types across Predicates in
the English Durative Construction” the relationship between aspectual
composition and coercion. He uses Collostructional Analysis (Ste-
fanowitsch and Gries 2003) to investigate the English durative construc-
tion. While the results show that the construction is significantly associated
with atelic verbs, as many as 25% of the data exhibit coercion effects,
where the construction causes generally telic verbs to become atelic. His
study therefore supports a constructional perspective on meaning construal
in which both the verb and the construction contribute meaning aspects and
illustrates how quantitative methods can contribute to shedding light into
compositional semantics.
In “The Force Dynamics of English Complement Clauses: A Usage-
based Account”, Martin Hilpert investigates English gerund classes with
infinitive complements using a collexeme (Stefanowitsch and Gries 2003)
analysis to identify the lexical material associated with this construction.
The results from the quantitative study are then investigated in terms of
Talmy’s (2000) force dynamics. The quantitative evidence supports the
qualitative analysis showing that the construction under consideration is
 Quantitative Methods in Cognitive Semantics 53

primarily used to express force dynamic meanings and that judgements on

the well-formedness of novel examples can be predicted on the basis of the
model developed.
In “Accounting for the Role of Situation in Language Use in a Cognitive
Semantic Representation of Sentence Mood”, Kerstin Fischer addresses
how both quantitative and qualitative methods can be used to establish
speakers’ different construals of an objectively identical situation. Situation
is a notoriously difficult concept in Cognitive Semantics because Cognitive
Semantics assumes categories to be subjectively construed by language; a
quantitative approach may provide the missing link since quantitative
methods may correlate the differences in subjective construal with different
sets of linguistic choices. Fischer illustrates her proposal on the choice of
grammatical mood in task-oriented interactions. The interdependencies
between situational construal and grammatical choice are then modelled in
a computational cognitive semantic representation of grammatical mood.
Arne Zeschel’s article “Exemplars and Analogy: Semantic Extension in
Constructional Networks” addresses the question of how speakers gener-
alise from conventional, learned expressions to a productive constructional
schema that allows the creation of new, unconventional instances. He in-
vestigates the question by considering the relationship between the number
of already established types per semantic reading and novel instances in an
increasingly more fine-grained semantic analysis. He demonstrates a pro-
cedure that takes great care to avoid pitfalls like those discussed in the pa-
pers by Newman and Schmid. He finds for the three semantic levels of
analysis large positive correlations between the semantic type frequency of
established uses and the number of creative extensions, allowing conclu-
sions regarding schematization based on semantic types.
The next section concerns specific methodological questions. It is lead
by Stefanie Wulff’s paper “Marrying cognitive-linguistic and corpus-based
methods: On the compositionality of English V NP-idioms” which focuses
on a core semantic issue, compositionality. Cognitive linguistics suggests
that syntactic structures be signs similar to lexical items, combining form
and meaning directly. In such an approach, both lexical items and construc-
tions contribute to the meaning of the whole construct, and they may do so
to different degrees. Stefanie Wulff now develops a quantitative measure
based on a large scale corpus linguistic investigation that takes all of the
component words of a construction into consideration. This measure re-
flects differences is entrenchement of instantiations of the V NP-
54 Kerstin Fischer

construction (e.g. make the headway, call the police, tell the story), using
the number of shared collocates as a measure for semantic similarity.
The following paper illustrates the use of quantitative methods for the
study of lexical semantics. Dylan Glynn examines the polysemy of the
English verb bother and offers an example of multifactorial research cou-
pled with fine-grained semantic analysis. He focuses on one of the basic
hurdles of corpus-driven Cognitive Semantics – the subjective nature of
manual semantic coding. He argues that even highly subjective semantic
coding can be verified using confirmatory statistical analysis. Although
objectivity may be impossible with this approach, verification may be the
key to enabling the field to advance in this direction. Glynn argues that
statistical modelling may further verify the results.
In “Beyond the Dative Alternation: The Semantics of the Dutch aan-
Dative”, Timothy Colleman addresses the Dutch prepositional dative con-
struction from a cognitive semantic perspective, thus focusing not on its
distinctive features that distinguish its meaning from the near-synonymous
Double Object Construction, but instead providing a semantic description
of the conceptual meaning of the construction under consideration. He uses
a Collexeme Analysis (Stefanowitsch and Gries 2003) to identify the verb
attracted to the construction under consideration. The ranking reveals a
broad semantic range in which possessional transfer plays a major role –
however, by far not the only one. He therefore introduces a multidimen-
sional analysis (cf. Geeraerts 1998) which starts with the definition of the
semantic core of the construction and proceeds by identifying metaphori-
cally motivated sense extensions along certain dimensions, in this case, the
two dimensions contact and direction, resulting in a family of ‘caused con-
tact’ senses.
In her article “Corpus-based Evidence for an Idiosyncratic Aspect-
Modality Relation in Russian”, Dagmar Divjak discusses the relationship
between aspect and modality in Russian, evaluating two competing theo-
retical proposals; her corpus linguistic analysis not only supports one of the
competing models empirically, but is also in accordance with cognitive
semantic analyses of the Russian aspectual system. She thus shows how
quantitative analyses outperform intuition-based studies and provides the
necessary correction of previous hypotheses in linguistic typology, support-
ing careful hand-crafted studies and tying in with cognitive models of as-
pect in Russian.
The volume is concluded with two papers that address the role of quan-
titative methods in Cognitive Semantics in general. Stefan Gries and
 Quantitative Methods in Cognitive Semantics 55

Dagmar Divjak take up criticisms and reservations against the use of

quantitative methods in Cognitive Linguistics. After addressing each of
these reservation, they outline the behavioural profile approach in their
chapter “Quantitative Approaches in Usage-based Cognitive Semantics:
Myths, Erroneous Assumptions, and a Proposal” to illustrate its use for the
investigation of two of the core problems in Cognitive Semantics: polys-
emy and the identification of prototypical word senses. A behavioural pro-
file consists of the percentages of (co-)occurrence of a word or sense with
respect to ID tags, that is, certain morphological, syntactic and semantic
properties of a word and a context. The authors continue by showing how
such a profile can be used by means of a cluster analytic approach to ad-
dress polysemy, near-synonymy and contrastive semantics. The results of a
behavioural profile approach can moreover be compared with experimental
evidence.
Finally, Anatol Stefanowitsch discusses in “Empirical Cognitive Se-
mantics: Some Thoughts” the development of Cognitive Semantics ‘from
an art to a science’, in which the use of empirical methodologies plays a
key role. He introduces the concept of operationalization and illustrates
how this concept can be applied to research in Cognitive Semantics focus-
ing on one of four kinds of meaning: meaning as concept, proposition, re-
ference and use. He then considers what consequences the empirical ap-
proach resting on the operationalization of meaning may have on the
development of Cognitive Semantics and concludes that this move will
change the discipline radically while providing Cognitive Semantics with
an empirically grounded, scientific basis.
The research combined in this volume presents innovative approaches to
genuinely semantic problems by means of quantitative methods, some of
which developed especially for the current purposes, others applied from
other research areas to the study of semantic phenomena. The volume also
provides an overview of different quantitative methodologies for the study
of Cognitive Semantics (for example, collexeme, correlation and regression
analysis) and a discussion of the theoretical and methodological conse-
quences of the quantitative approaches taken.
56 Kerstin Fischer

Notes

1. Some scholars understand this criterion to be defining for scientific studies in

general (see, for instance, Stefanowitsch, this vol.).

References

Bednarek, Monika
2008 Semantic preference and semantic prosody re-examined. Corpus
Linguistics and Linguistic Theory 4(2): 119-139.
Berlin, Brent and Paul Kay
1969 Basic Color Terms. Their universality and evolution. Berkeley, CA:
University of California Press.
Elman, Jeff
2006 Computational approaches to language acquisition. In Encyclopedia
of Language & Linguistics, K. Brown (ed.), 726-732. 2nd ed., Vol.
2. Oxford: Elsevier.
Fillmore, Charles J.
1975 An alternative to checklist theories of meaning. Proceedings of the
Berkeley Linguistics Society 1: 123-131.
1976 Frame semantics and the nature of language. Annals of the New York
Academy of Sciences: Conference on the Origin and Development of
Language and Speech 280: 20-32.
1982 Frame semantics. In Linguistics in the Morning Calm, Linguistic
Society of Korea (ed.), 111-137. Seoul: Hanshin Publishing.
1988 The mechanisms of ‘Construction Grammar’. Proceedings of the
Fourteenth Annual Meeting of the Berkeley Linguistic Society 14:
35-33.
Fillmore, Charles J., Paul Kay and Mary O’Connor
1988 Regularity and idiomaticity in grammatical constructions: The case
of let alone. Language 64(3): 501–538.
Fillmore, Charles J. and Beryl T. Atkins
1992 Towards a frame-based lexicon: The case of RISK. In Frames,
Fields, and Contrasts: New essays in semantic and lexical organiza-
tion, Adrienne Lehrer and Eva Kittay (eds.), 75–102. Mahwah, N.J.:
Erlbaum Publishers.
 Quantitative Methods in Cognitive Semantics 57

Fischer, Kerstin
2000 From Cognitive Semantics to Lexical Pragmatics: The functional
polysemy of discourse particles. Berlin / New York: Mouton de
Gruyter.
Geeraerts, Dirk
1998 The semantic structure of the indirect object in Dutch. In The Dative.
Vol. 2. Theoretical approaches, Willy Van Langendonck and Wil-
liam Van Belle (eds.), 185-210. Amsterdam / Philadelphia: John
Benjamins.
2007 Methodology in Cognitive Linguistics. In Methods in Cognitive
Linguistics, Mónica González-Márquez, Irene Mittelberg, Seana
Coulson and Michael J. Spivey (eds.), 21-50. Amsterdam / Philadel-
phia: John Benjamins.
Geeraerts, Dirk, Stefan Grondelaers and Peter Bakema
1994 The Structure of Lexical Variation. Meaning, naming, and context.
Berlin / New York: Mouton de Gruyter.
Gibbs, Raymond W. Jr.
2007 Why Cognitive Linguistics should care more about empirical meth-
ods. In Methods in Cognitive Linguistics, González-Márquez,
Mónica, Irene Mittelberg, Seana Coulson and Michael J. Spivey
(eds.), 27-54. Amsterdam / Philadelphia: John Benjamins.
Glynn, Dylan
2009 Polysemy, syntax and variation. In New Directions in Cognitive
Linguistics, Vyvyan Evans and Stephanie S. Pourcel (eds.), 77-104.
Amsterdam / Philadelphia: John Benjamins.
González-Márquez, Mónica, Irene Mittelberg, Seana Coulson and Michael J.
Spivey (eds.)
2007 Methods in Cognitive Linguistics. Amsterdam / Philadelphia: John
Benjamins.
Gries, Stefan and Anatol Stefanowitsch
2004 Extending collostructional analysis: A corpus-based perspective on
‘alternations’. International Journal of Corpus Linguistics 8: 31-61.
Gries, Stefan Th. and Anatol Stefanowitsch (eds.)
2006 Corpora in Cognitive Linguistics. Corpus-based approaches to
syntax and lexis. Berlin / New York: Mouton de Gruyter.
Gries, Stefan and Dagmar Divjak
2009 Behavioral profiles: A corpus-based approach to Cognitive Semantic
analysis. In New Directions in Cognitive Linguistics, Vyvyan Evans
and Stephanie S. Pourcel (eds.), 57-75. Amsterdam / Philadelphia:
John Benjamins.
58 Kerstin Fischer

Gries, Stefan Th. and Anatol Stefanowitsch

2010 Cluster analysis and the identification of collexeme classes. In Em-
pirical and Experimental Methods in Cognitive/Functional Re-
search, Sally Rice and John Newman (eds.), 73-90. Stanford: CSLI.
Heritage, John
1995 Conversation analysis: Methodological aspects. In Aspects of Oral
Communication, Uta M. Quasthoff (ed.), 391-418. Berlin / New
York: Mouton de Gruyter.
Heylen, Kris, Yves Peirsman and Dirk Geeraerts
2008 Automatic synonymy extraction. A comparison of syntactic context
models. Computational Linguistics in the Netherlands 2007, Suzan
Verberne, Hans van Halteren, Peter-Arno Coppen (eds.), 101-116.
Utrecht: LOT.
Kristiansen, Gitte, Michel Achard, René Dirven and Francisco José Ruiz de Men-
doza Ibáñez (eds.)
2006 Cognitive Linguistics: Current applications and future perspectives.
Berlin / New York: Mouton de Gruyter.
Lakoff, George
1987 Women, Fire, and Dangerous Things. What categories reveal about
the mind. Chicago: University of Chicago Press.
Langacker, Ronald W.
1987 Foundations of Cognitive Grammar, Vol. 1. Theoretical prerequi-
sites. Stanford: Stanford University Press.
1991 Concept, Image, Symbol. Berlin / New York: Mouton de Gruyter.
2000 A dynamic usage-based model. In Usage-Based Models of Lan-
guage, Michael Barlow and Suzanne Kemmer (eds.), 1-63. Stanford:
CSLI.
2008 Cognitive Grammar: A basic introduction. Oxford: Oxford Univer-
sity Press.
Stefanowitsch, Anatol
2009 Cognitive Linguistics as a Cognitive Science. Keynote Lecture at the
Conference on Bi-directional Perspectives in the Cognitive Sciences
in Marburg, Germany, 2nd March, 2009, available at http://www-
user.uni-bremen.de/~anatol/.
Stefanowitsch, Anatol and Stefan Th. Gries
2003 Collostructions: Investigating the interaction of words and construc-
tions. Proceedings of the 18th International Conference on Compu-
tational Linguistics, 747-753.
2005 Co-varying collexemes. Corpus Linguistics and Linguistic Theory
1(1): 1-46.
2008 Channel and constructional meaning: A collostructional case study.
In Cognitive Sociolinguistics: Language variation, cultural models,
 Quantitative Methods in Cognitive Semantics 59

social systems, Gitte Kristiansen and René Dirven (eds.), 129–152.

Berlin / New York: Mouton de Gruyter.
Tummers, Jose, Kris Heylen and Dirk Geeraerts
2005 Usage-based approaches in Cognitive Linguistics: A technical state
of the art. Corpus Linguistics and Linguistic Theory 1(2): 225-261.
Section I

Corpus methods in Cognitive Semantics

The doctor and the semantician

Dirk Geeraerts

Abstract
Starting from an analogy with medical reasoning, the paper argues for the import-
ance of an empirical method in linguistics, and more specifically, in semantics.
Against the position that an introspection-based semantic methodology is suffi-
cient, it argues that such a position hinges on two debatable assumptions (the as-
sumption of a native speaker’s complete command of the language, and the as-
sumption of direct access to meaning). The counterposition taken in the paper is a
nuanced one, however: the fact that an introspection-based semantic method is not
sufficient does not imply that intuition in semantics is superfluous. In the second
half of the paper, a systematic analysis of the concept of an empirical cycle forms
the basis for an appraisal of the role and relevance of interpretation in semantic
research. It is suggested that the empirical method in semantics, rather than being
the opposite of a more traditional hermeneutic approach, broadens the basis of the
hermeneutic process with maximally objectifiable data and shores the semanti-
cian’s interpretative efforts with quantitative techniques.
Keywords: methodology, introspection, interpretation, hermeneutics, empirical
cycle, linguistic evidence

1. Linguists in the surgery

Suppose you have a patient who has been suffering for a number of days
from a muscular ache in the upper legs, regularly accompanied by a gnaw-
ing pain in the right groin. The night before you see him, he was kept
awake by sharp pains in the lower back, in the region of the sacral spine.
Several interpretations suggest themselves: an abdominal or inguinal rup-
ture, a spinal hernia, an inflammation – each time accompanied by a radiat-
ing of the pain from the initial location to other parts of the body. The
possibility of an abdominal or inguinal rupture is easily discarded: there is
no visible swelling, and neither palpation nor the valsalva maneuver reveals
any. Given the fact that the patient’s movements are unimpeded, and that
the lasègue test is negative, a spinal hernia or similar is not very likely
either, but to make sure, you have a number of x-rays of the pelvis and the
64 Dirk Geeraerts

lower spine taken. They confirm that the option of a spinal hernia may be
rejected. Remains the possibility of an inflammation, but a blood test does
not reveal any heightened values of C-reactive protein. However, a CT-
scan of the lumbal spinal column displays a diffuse picture for the sacro-
iliac joints, so that you can now safely conclude to bilateral sacroiliitis. As
a therapy for the inflammation of the sacroiliac joints, you suggest anti-
inflammatory drugs.
What you do here, as a medical doctor, is putting into practice an em-
pirical method of investigation: instead of just relying on your intuition,
you think about the symptoms that would correlate with one or the other
interpretation that you come up with, and then check whether those corre-
lates indeed obtain. Your overall interpretative hypothesis is analyzed into
separate, objectively testable correlates, and the best correlation yields the
best interpretation. You look for convergence among different types of
evidence (physical tests, blood tests, medical imaging), and the evidence
that you take into consideration is not chosen randomly: the fact that a cer-
tain type of evidence correlates with a certain disease or disorder is itself
based on a large number of observations, taken from numerous patients
throughout the history of medicine.
Compare this to what you traditionally do as a semanticist: the proced-
ures you follow are much more intuitive, and while both the doctor and the
semantician may start from hunches, going beyond those hunches in a sys-
tematic and controllable way is much less common in semantics than in
medicine – notwithstanding the recent upsurge of quantitative studies of the
kind collected in this volume. Why is that? What are the factors that pre-
vent linguistic semantics from embracing the empirical method more fre-
quently and enthusiastically than it actually does? Why is a volume like the
present one still relatively exceptional in the study of natural language
meaning?
One alternative to consider could be that it is impossible to treat seman-
tics in this way. Semantics, of all disciplines, would seem to be most resis-
tent to quantification: how can meaning, the most qualitative of all lin-
guistic features, be expressed in numbers, and more broadly, how could
meaning, the most ephemeral and subjective of all linguistic phenomena, be
tackled with methods that aim at objectivity? The suggestion of an impos-
sibility may however be rejected on two grounds.
First, with a rhetorical repartee, if psychology can do it, why couldn’t
linguistics? If we believe that the methods of experimental psychology –
controlled experimentation and quantitative analysis of the data – constitute
 The doctor and the semantician 65

a legitimate and fruitful way of investigating cognitive phenomena at large,

then surely linguistic meaning may be subjected to the same kind of ap-
proach – as psycholinguists already do, of course.
Second, if we think of the traditional method of lexical semantics, we
may definitely discern the basics of the empirical approach as outlined by
our medical example. Think, for instance, of a historical linguist determin-
ing the reading of a passage in a historical text. The interpretative process,
like the physician’s interpretative effort, is guided by contextual clues (tex-
tual ones, and extra-textual, historical or cultural ones), and in general –
disregarding deliberately recalcitrant hermeneutic theories like the decon-
structivist approach – those interpretations will be favoured that make sense
of most of the clues. From Boeckh (1877) to Joos’s brilliant ‘The best
meaning is the least meaning’ (1972), the principles of such a coherence-
based method of interpretation have been spelled out fairly explicitly. Simi-
larly, looking for generalizations has just as well been part and parcel of
semantics. What else, for instance, is the classical interest in metaphor and
metonymy but an attempt to establish regularities of semantic change? If
we think of these well-established facets of semantics, the transition to a
quantitative methodology would seem to be a gradual step, a continuation
rather than a rupture – an observation that makes it even less probable that
the adoption of an empirical methodology is an impossibility.

2. The dangers of self-diagnosis

So, given that there is little credibility to the claim that semantics is impos-
sible to treat in this way, maybe there is a widespread assumption that it is
not necessary to resort to empirical methods of the type described above.
Perhaps the extra trouble of moving beyond a largely intuitive approach is
not worth it: perhaps the more intuitive approach is sufficient, in the sense
that it allows us to reach the same conclusions as a more empirical one.
Given this assumption, how can we argue for the inevitability of adopt-
ing the scientific method in linguistics? A rhetorical appeal to the ‘univer-
sally accepted’ nature of the scientific method will not do, for two closely
related reasons. First, applying the very principle of falsifiability makes it
easy to see that the statement about the universal acceptedness of the scien-
tific method is in fact not universally valid: the method is not universally
accepted, at least not in the humanities. Competing methods like herme-
neutics and phenomenology, to name only these, do exist, and their exist-
66 Dirk Geeraerts

ence suffices to show that the appeal to the ‘universally accepted’ nature of
the scientific method is merely a prescriptive strategem. Second, it could be
argued that the object of study in linguistics is sufficiently different from
that of the natural sciences to warrant a radically different method. This, of
course, is a time-honoured line of argumentation, dating back to Dilthey
(1910): if there is an essential distinction between the natural sciences and
the human sciences, then linguistics too will be seen as a Geisteswissen-
schaft rather than a Naturwissenschaft, and it will be pursued with appro-
priate interpretative methods.
Let us therefore rather take an analytic approach, and consider the ques-
tion under which conditions the introspective method would be sufficient.
To bring some system into the discussion, let us note that studying a phe-
nomenon involves demarcating the object of enquiry, plus observing that
object from a certain perspective. If you study the social monogamy of
coyotes, you need to distinguish coyotes from jackals, and you need to
observe whether coyotes have a life-time social bond with one partner or
not (regardless of the presence of extra-pair copulation). Similarly, if you
intend to study linguistic meaning, you need to distinguish what expres-
sions belong to the investigated language(s), and you will need to make
observations about the meaning of those expressions. For both aspects, the
demarcational one and the observational one, we may now ask the question
why the scientific method is superior over introspection – or perhaps more
cautiously, why the introspective method is not sufficient. If we try to iden-
tify the conditions that would make the intuitive method sufficient, what
evidence do we have to claim that those conditions are met with or not?
The argument pro intuition would seem to hinge on two assumptions:
the assumption of a native speaker’s complete command of the language,
and the assumption of direct access to meaning. Let us consider both as-
sumptions in turn.

1. The demarcation of a language boils down to the identification of the

expressions of the language. Introspection will do the job if, as a competent
speaker of the language, we have complete command of the language. This
is Chomsky’s ‘ideal speaker/hearer’, where he said that
Linguistic theory is primarily about the language of an ideal speaker-hearer,
in a completely homogeneous speech community, who knows its language
perfectly and is unaffected by grammatically irrelevant conditions, such as
memory limitations, distractions, shifts of attention and interest, and errors
 The doctor and the semantician 67

(random or characteristic) in applying his knowledge of the language in ac-

tual performance (Chomsky 1965: 3).
If you are an ideal speaker-hearer of this type, introspection is sufficient,
but of course, the very notion of such an ideal speaker-hearer is too much
of an idealization to be valid as the methodological basis of linguistics.
Languages are not homogeneous, and competent speaker-hearers of the
language do not have a full command of the language, not even if we re-
strict the perspective to passive, receptive knowledge: no individual
speaker has a complete command of the language. Consider vocabulary
knowledge: who could claim to know all the words of the language? If we
demarcate the lexicon of the language, for instance by using a reference
dictionary, we would have difficulty finding speakers with an active (or
even passive) knowledge of the whole vocabulary – perhaps not even the
lexicographers compiling the dictionary would pass the test. The argument
against homogeneity has been spelled out many times, as have the method-
ological consequences: see Labov (1971) for a cogent refutation of the
Chomskyan stance, and compare Geeraerts (2005) for a discussion of It-
konen’s (2003) non-Chomskyan variant of the argument in favour of intro-
spection. (On good grounds, Itkonen distinguishes ‘intuition’ from ‘intro-
spection’, but the methodological thrust of Itkonen’s reasoning is the
same.)
However, we need to take a further step by considering the possibility of
a ‘local’ interpretation of homogeneity. We may not know the language –
the full spectrum of linguistic behaviour exhibited by different subgroups
of a linguistic community in different contexts – as such, but we may still
be able to demarcate our own language perfectly well. We may not know
for sure what others say, the reasoning goes, but we definitely know what
we would say in specific circumstances. The methodological value of intro-
spection would be salvaged then: if each of us separately can give a reli-
able, first-hand report on his or her language behaviour, we can generalize
over those reports to obtain an adequate image of the overall heterogeneity,
without recourse to objective data-gathering methods of the type that are
usual in sociolinguistics.
But self-reports are known to be unreliable. Although they may not re-
ject the approach in all respects, classic studies like Nisbett & Wilson
(1977), Ericsson & Simon (1980), Webb, Campbell, Schwartz & Seechrist
(1981) indicate the limits on the validity of self-reports in the cognitive
sciences, and even in disciplines that use self-report extensively, like some
branches of social science and medical studies, the value of the method is a
68 Dirk Geeraerts

cause of constant concern: studies comparing subjective reports and objec-

tive measurements are conducted to check the validity of the method. In
linguistics, one need only consider the quandaries of grammaticality
judgements to appreciate the restricted value of an introspective demarca-
tion of one’s own idiolect. We have probably all had the experience that in
assigning grammaticality judgements – on the basis of our own idiolect –
we would soon wind up with numerous question marks: even if we stay
within our own idiolect, we are not entirely certain what actual linguistic
behaviour we exhibit, and our intuitions about grammaticality tend to ev-
aporate under our own eyes. With regard to this point too, the argument has
been made forcefully elsewhere: see Schütze (2006) and Sampson (2007)
for a very thorough recent discussion. Overall, we may conclude with
Sampson that there is no reason to assume that patterns in a speaker’s intui-
tive grammaticality judgments reflect realities of his language.

2. With regard to the observational part of the methodology, introspection

will be vindicated if we assume that meaning is a directly observable prop-
erty of linguistic expressions. Note that this argument in favour of intro-
spection is independent of the previous, demarcational point. Even if we
accept that we need to have recourse to objective data-gathering to get an
idea of what actually occurs in the language (rather than what we think
occurs), we could still maintain that our understanding of each of the ex-
pressions so encountered is immediate and indubitable. In general, the em-
pirical method ss based on the notion that every descriptive term in the
language of science needs to be connected with terms designating observ-
able properties of things. Further, this connection should be of such a kind
that statements applying such terms can be intersubjectively confirmed.
Given such a description, meaning would seem to be an observable prop-
erty of linguistic expressions. After all, when we read The glass is on the
table, we observe something more than a mere succession of sounds: we
observe something conceptual, notional, cognitive; we know that some-
thing is meant, and we think we know what is meant.
Now surely, it makes no sense denying that we understand language,
and that that understanding can be part of our methodology. But how far
does that understanding go, and does it go far enough to halt the incorpora-
tion of the scientific method into semantics? The problems we may identify
mirror the ones we mentioned with regard to the demarcational use of
introspection.
 The doctor and the semantician 69

To begin with, consider the issue of homogeneity. We cannot simply as-

sume that our own reading of an utterance, however clear it may seem to
us, reflects a common interpretation: we would need to go beyond our own
observations to reach certainty on this point. But is this really a problem?
As with regard to the demarcational issue, the global problem would not be
decisive if a local transparancy is maintained: individually, we may have no
direct access to other people’s interpretations, but as long as we have direct
access to our own meanings, a picture of the global situation may be ob-
tained indirectly, through comparison and discussion. If subjectively we
have direct access to our own interpretations, we can intersubjectively
compare interpretations to see if we understand language in the same way.
One philosophical problem with this position is the danger of an infinite
regress. To compare our individual interpretations, we would need a lan-
guage, and the language would have to be a common one, i.e. one that we
know is the same for all of us. How else could we compare our individual
interpretations in a safe way? So, in order to establish whether we are
speaking a given language in the same way, we need to establish whether
we are speaking our language of intersubjective comparison in the same
way – and so on ad infinitum. It is not very likely that we could get out of
this regress if we don’t find a way of methodologically transcending our
individual intuitions, by correlating them with more objectifiable observa-
tions.
Further, whether we have indubitable direct access to our own interpre-
tations is as doubtful as the idea that we have a complete introspective
overview of our own speech habits. Take the simple example we mentioned
above: The glass is on the table. What would it mean to say that we clearly
and distinctly understand such an utterance? When we say the glass is on
the table, we understand, among other things, that there is an object that
belongs to the category ‘glass’. If we claim that our understanding of the
utterance is perfectly transparent to ourselves, that would mean that in the
process of understanding the utterance, we realize a clear and distinct idea
of the category ‘glass’. Such a clear and distinct idea of ‘glass’ would in-
clude an answer to a number of issues that typically relate to natural lan-
guage categorization: what are the boundaries of the category, both inten-
sionally and extensionally? Is shape the defining characteristic, so that
plastic or earthenware glass-shaped objects would still be glasses? Or is it
function? Or is it material? Or specific combinations of these features? If
our direct access to meaning would be as transparent as the intuitionists
claim, we would have an immediate, mentally accessible answer to those
70 Dirk Geeraerts

questions when we understand an utterance like The glass is on the table.

But clearly, we do not. (This point has been made before: see the discussion
in Geeraerts 1999, Kristiansen and Geeraerts 2007 of Wierzbicka’s Natural
Semantic Metalanguage.)
The point may be further illustrated with an example from recent re-
search in Cognitive Linguistics. Kemmer & Verhagen (1994) have argued
that the choice between the causative auxiliaries doen and laten in Dutch is
determined by (in)direct causation. Indirect causation refers to ‘a situation
that is conceptualized in such a way that it is recognized that some other
force besides the initiator is the most immediate source of energy in the
effected event’ (Verhagen & Kemmer 1997: 6). Typically, in a sentence
like De agent liet de studenten passeren ‘The police man let the students
pass’, the students ultimately do the passing: the policeman only creates the
conditions for the students to perform the action. Now, if the idealist view
of direct access to meaning were correct, the presence of (in)direct causa-
tion would be easy to establish: we would just have to inspect our own
interpretation of utterances like De agent liet de studenten passeren. If we
want to avoid circularity, that would mean being able to determine whether
sentences containing the causative construction exhibit direct or indirect
causation regardless of whether the causative verb is filled out by doen or
laten.
So how, in a sentence like The teacher CAUSE the students finish their
book, could (in)directness be established? As we probably have no direct
intuition of (in)direct causation, we will look for ways of operationalizing
the concepts in question. Like our doctor, we will look for indices of
(in)direct causation that are easier to establish than (in)direct causation as
such. Following d’Andrade (1987), Kemmer and Verhagen for instance
assume that we conceptualize situations with animate causers and causees
in such a way that animate beings are not normally thought of as acting
directly upon other human beings. While physical entities and forces are
taken to exert a direct action on other things, animate beings exert an influ-
ence on others only indirectly, through the intervening medium of the phys-
ical world. It follows that direct causation is considered typical for physical,
inanimate contexts, whereas animate contexts prime for indirect causation.
But we will probably have to take a few more steps (Speelman & Geer-
aerts 2010), and the operational criteria may well turn out to be fairly com-
plex, semantically speaking. For instance, while the description provided
by Kemmer and Verhagen suggests that the distinction between physical
and immaterial causation plays a role, a further analysis would have to re-
 The doctor and the semantician 71

fine that distinction. Difficult cases are likely to occur: if the lightning
made the alarm go off is material causation, should we then also classify
the lightning made the children tremble as physical? The latter sentence
probably features a less material type of causation than the former, but at
the same time, the lightning made the children tremble would seem to be
more material than the idea of having to stay alone at home made the chil-
dren tremble. So where would we draw the line? The methodological point,
however, would not be to enforce a binary categorial decision in every pos-
sible case, but rather to find an operationally applicable set of diagnostic
features that would make it possible to chart all possible borderline cases
and nuances. Such a componential analysis of the relevant contexts of use
would indeed almost inevitably imply that the concepts ‘direct’ and ‘indi-
rect causation’ stop being categorial variables, but rather reveal themselves
as prototypical reference points on a continuum (or perhaps even in a
multidimensional semantic space).
In general, then, the demarcational use of introspection rests on the as-
sumption of homogeneity in a linguistic community, and the observational
use of introspection rests on the assumption of direct access to meaning in
the individual. These are both assumptions that have to be taken seriously,
as both have a considerable pedigree in the history of the linguistics and the
human sciences: the former has its origins in the structuralist conception of
the linguistic system, and the latter in an idealist philosophy of mind. That
is to say, an indignant reaction of the type ‘How is it possible that at the
beginning of the 21st century we still have to argue for the adoption of
empirical methods in linguistics?’ is unnecessarily naive. It is true that
compared to psychology, linguistic semantics lags behind roughly a century
in the large-scale adoption of an empirical methodology, but the retardation
is not due to ignorance or inflexibility, but rather to deeply held assump-
tions that deserve to be examined in their own right. However, as both as-
sumptions appear to be questionable upon such an examination, we may
conclude that the intuitive method is not sufficient.

3. Evidence-based medicine – and linguistics too

But does that mean, conversely, that intuition can be dispensed with?
We’ve given a negative answer to the question whether an empirical meth-
odology might be unnecessary, but should we now conclude, turning the
question round, that intuition is not only not sufficient, but also not neces-
72 Dirk Geeraerts

sary? In order to answer that question, we should first try to summarize the
central aspects of research as meant by the scientific method, where the
‘scientific method’ - as in our good doctor’s endeavours - is the approach to
scientific investigation in which the empirical testing of hypotheses is
paramount: systematic data gathering on the basis of observation or ex-
perimentation yields material that may be used to falsify predictions de-
rived from a theoretical hypothesis. (This passage is an elaboration of a
number of remarks made in Geeraerts 2006.) What are the main features of
empirical research?
First, empirical research is data-driven. You cannot easily draw conclu-
sions from single cases and isolated observations, and the more data you
can collect to study a particular phenomenon, the better your conclusions
will get.
Second, empirical research in linguistics may be observational or ex-
perimental; there is a complementarity between both approaches. The re-
search data may come from different sources: they may be collected as they
exist (as is the case in corpus research), but they may also be elicited by
doing experimental research, or by doing survey research. As applied to
language, the mutual advantages of observational versus experimental re-
search are clear: observational research (viz. corpus research) allows you to
study language in a natural and spontaneous state; but experimental re-
search, by contrast, may give you a better control over specific variables, as
when they are underrepresented in the corpus.
Third, empirical research involves quantitative methods. In order to get
a good grip on the broad observational basis of elicited and/or non-elicited
data, investigators need techniques to come to terms with the amount of
material involved. Specifically, they will need statistical tests to determine
whether specific observations might be due to chance or not. It needs to be
emphasized that quantification is not the essence of empirical research, but
simply follows in a natural way of what an empirical methodology tries to
achieve: quantification in empirical research is not about quantification, but
about data management and hypothesis testing.
Fourth, empirical research crucially hinges on asking the right ques-
tions, or in other words, on the formulation of hypotheses. No perception
could be more misguided than to think that once you have your database of
elicited or non-elicited observations, the conclusions will arise automati-
cally and purely inductively from the data. On the contrary, the only con-
clusions you will be able to draw are the ones that relate to hypotheses you
have formulated and tested – so that will be the investigator’s first task.
 The doctor and the semantician 73

Another way of saying this is that empirical research necessarily combines

inductive and deductive reasoning: on the one hand, you work in a bottom-
up way from data to hypotheses, but on the other hand, those hypotheses
will also be derived top-down from the theoretical perspective you adopt in
thinking about your data.
Fifth, empirical research requires the operationalization of hypotheses.
It is not sufficient to think up a plausible and intriguing hypothesis: you
also have to formulate it in such a way that it can be put to the test. That is
what is meant by ‘operationalization’: turning a hypothesis into concrete
predictions that can be tested against the data. In most empirical research in
linguistics, it is questions of operationalization that require all the ingenuity
of the researcher – and most of his or her time, because getting the relevant
data and measurements is not an automatic process.
And sixth, empirical research involves an empirical cycle in which sev-
eral rounds of data gathering, testing of hypotheses, and interpretation of
the results follow each other. Just like it is misguided to think that empiri-
cal, data-driven research automatically gives one all the answers, it is mis-
guided to think that it immediately gives one the final answer. The empiri-
cal cycle as such, in fact, does not constitute a straightforward march
towards the truth, because negative results may be interpreted in different
ways. If a prediction is not borne out, at least two kinds of interpretation
suggest themselves: the original hypothesis (or the broader framework in
which it is couched) may be wrong, but in principle, it could also be the
case that our operationalization of the hypothesis was not adequate. The
assumption may be wrong, or our way of testing the assumption may be
inappropriate - but the consequences in either case are largely different.
Empirical research seeks maximal objectivity, but it is in no way a me-
chanical procedure that inevitably leads to a single possible result. That is
not the way it happens in the hard sciences, and it is not the way it happens
in the study of language either.
This is a point that cannot be sufficiently emphasized, because the way
in which the empirical cycle is often represented in introductory treatises on
scientific method is often overly optimistic. The stereotyped view, so to
speak, is represented in Figure 1. A theory leads to predictions in the form
of specific hypotheses, which are experimentally (or more broadly, oper-
ationally) tested. The analysis leads back to the theory: the experimental
results may or may not contradict the theoretical expectations.
74 Dirk Geeraerts

Figure 1. The empirical cycle in its standard form

In actual practice, however, negative experimental results are not necessa-

rily seen as straightforward falsifications of theories. For one thing, nega-
tive results may be attributed to inappropriate experimental designs rather
than to theoretical mistakes. The famous Michelson and Morley experi-
ments are a case in point: instead of rejecting the assumption that there is
an ether (which would seem obvious from our contemporary point of
view), their failure to measure an ether wind only stimulated them to find
ever more sophisticated experimental designs to observe the ether wind.
Further, rather than to a wholesale rejection of a theory, negative results
may lead to modifications in the theory and refinements of the hypotheses.
This, after all, is the basis of the Kuhnian theory of science: as against the
strict falsificationalist model of scientific enquiry proposed by Popper
(1963), Kuhn (1970) and his followers (Lakatos 1974) stress the fact that
scientific paradigms have a tendency towards inertia. Rather than rejecting
theories on the basis of single negative results, theories are adapted to deal
with the critical findings; and (apart from the specific dynamics of the soci-
ology of science) it is basically only when the adaptive capacity of a theory
runs that it may be abandoned as a whole.
 The doctor and the semantician 75

Figure 2. The empirical cycle in a sophisticated version

In Figure 2, a more sophisticated version of the empirical cycle is drawn:

upon analysis, experimental results may lead (in decreasing order of theo-
retical radicality) to a change of theory, a change of theoretical assumptions
and hypotheses within a theory, or a change of experimental design. From
the point of view of the present discussion, the non-mechanical nature of
scientific progress again testifies to the importance of subjective elements
in empirical research: results need to be interpreted and decisions need to
be made on the basis of those interpretations.

4. Interpretation, clinical and linguistic

What then about the role of intuition in empirical research? On the basis of
this nutshelled presentation of an empirical approach, we may conclude that
empirical research does not rule out creativity and intuition. To the undis-
cerning eye, the ideal of scientific objectivity would seem to banish the
investigator as a subject from the investigation, but a closer look makes
clear that ingenuity and interpretative insight are indispensable features of
the empirical cycle. Hypotheses translate an intuitive understanding into
operational predictions; finding the right operationalization rests on inven-
tiveness as much as on expertise; and processing the results of the empirical
cycle requires creative imagination. Empirical research does not lower the
76 Dirk Geeraerts

demands on the subjective skills of the researchers; it only raises the cri-
teria for the objective validity of their claims.
So will the doctor save the semantician? Or is the adoption of an em-
pirical method a failproof way of making the human scientist lose his soul?
It should be clear from the foregoing discussion that an empirical method
should not be treated fetishistically – neither as a positive nor a negative
fetish. It is not a mechanical march to absolute truth, but an inevitable and
indispensable next step for anyone dealing with the intricacy and elusive-
ness of natural language meaning: in a systematic attempt to overcome the
insufficiencies of an intuitive method, it broadens the basis of the herme-
neutic process with maximally objectifiable data and it supports the inter-
pretative efforts with quantitative techniques. Rather than being the oppos-
ite of a more traditional hermeneutic approach, an empirical approach to
semantics is the completion and consummation of it.

References

Boeckh, August
1877 Enzyklopädie und Methodenlehre der philologischen Wissen-
schaften. Leipzig: Teubner.
Chomsky, Noam
1965 Aspects of the Theory of Syntax. Cambridge, MA: MIT Press.
d’Andrade, Roy
1987 A folk model of the mind. In Cultural Models in Language and
Thought, Dorothy Holland and Naomi Quinn (eds.), 112-148. Cam-
bridge: Cambridge University Press.
Dilthey, Wilhelm
1910 Der Aufbau der geschichtlichen Welt in den Geisteswissenschaften.
Berlin: Verlag der Königlichen Akademie der Wissenschaften.
Ericsson, K. Anders and Herbert A. Simon
1980 Verbal reports as data. Psychological Review 87: 215-251.
Geeraerts, Dirk
1999 Idealist and empiricist tendencies in Cognitive Linguistics. In Cogni-
tive Linguistics: Foundations, scope, and methodology, Theo Jans-
sen and Gisela Redeker (eds.), 163-194. Berlin / New York: Mouton
de Gruyter.
2005 Lectal data and empirical variation in Cognitive Linguistics. In Cog-
nitive Linguistics. Internal dynamics and interdisciplinary interac-
 The doctor and the semantician 77

tions, Francisco José Ruiz de Mendoza Ibáñez and Sandra Peña

Cervel (eds.), 163-189. Berlin / New York: Mouton de Gruyter.
2006 Methodology in Cognitive Linguistics. In Cognitive Linguistics:
Current applications and future perspectives, Gitte Kristiansen,
Michel Achard, René Dirven and Francisco José Ruiz de Mendoza
Ibañez (eds.), 21-49. Berlin / New York: Mouton de Gruyter.
Itkonen, Esa
2003 What is Language? A Study in the Philosophy of Linguistics. Turku:
Åbo Akademis tryckeri.
Joos, Martin
1972 Semantic axiom number One. Language 48: 257-265.
Kemmer, Suzanne and Arie Verhagen
1994 The grammar of causatives and the conceptual structure of events.
Cognitive Linguistics 5: 115-156.
Kristiansen, Gitte and Dirk Geeraerts
2007 On non-reductionist intercultural pragmatics and methodological
procedure. In Explorations in Pragmatics. Linguistic, cognitive and
intercultural aspects, Istvan Kecskes and Laurence R. Horn (eds.),
257-286. Berlin / New York: Mouton de Gruyter.
Kuhn, Thomas
1970 The Structure of Scientific Revolutions. Chicago: Chicago University
Press.
Labov, William
1971 Some principles of linguistic methodology. Language in Society 1:
97-120.
Lakatos, Imre
1974 Falsification and the methodology of scientific research program-
mes. In Criticism and the Growth of Knowledge, Imre Lakatos and
Alan Musgrave (eds.), 91-196. Cambridge: Cambridge University
Press.
Nisbett, Richard and Timothy Wilson
1977 Telling more than we can know. Verbal reports on mental processes.
Psychological Review 77: 231-259.
Popper, Karl
1963 Conjectures and Refutations. The growth of scientific knowledge.
London: Routledge.
Sampson, Geoffrey
2007 Grammar without grammaticality. Corpus Linguistics and Linguistic
Theory 3: 1-32.
Schütze, Carston T.
2006 The Empirical Base of Linguistics: Grammaticality judgments and
linguistic methodology. Chicago: Chicago Unversity Press.
78 Dirk Geeraerts

Speelman, Dirk and Dirk Geeraerts

2010 Causes for causatives: The case of Dutch ‘doen’ and ‘laten’. In Cau-
sal Categories in Discourse and Cognition, Ted Sanders and Eve
Sweetser (eds.), 173-204. Berlin / New York: Mouton de Gruyter.
Verhagen, Arie and Suzanne Kemmer
1997 Interaction and causation: Causative constructions in modern stan-
dard Dutch. Journal of Pragmatics 27: 61-82.
Webb, Eugene J., Donald T. Campbell, Richard D. Schwartz and Lee Seechrist
1981 Unobtrusive Measures: Non-reactive research in the social sciences.
Chicago: Rand McNally.
Balancing acts:
Empirical pursuits in Cognitive Linguistics

John Newman

Abstract
Cognitive linguists, more so than linguists of other persuasion, are expected to be
open to a wide range of data and methodologies. Usage-based data and methodolo-
gies are playing an increasingly important role in Cognitive Linguistics and it is
appropriate to re-examine some of the common practices in dealing with this kind
of data. It is argued that a number of practices in the use of corpora and associated
methodologies should be evaluated more critically than they currently are. This
does not mean that we must reject current practices, but we should be more aware
of the need to weigh alternatives more carefully before we settle on one way of
proceeding.
Keywords: usage, corpus, spoken corpus, conversation, communication, part-of-
speech tags, quantitative methods, inflected form, lemma

1. Introduction

The terms Cognitive Linguistics and Cognitive Semantics are open to a

variety of interpretations, but here I will be guided by the characterizations
given in Evans and Green (2006). Broadly speaking, Cognitive Linguistics
is concerned with general principles that provide some explanation for all
aspects of language, including principles drawn from disciplines other than
linguistics (Evans and Green 2006: 27-28). If we accept this characteriza-
tion of the field, then many kinds of evidence and methodologies are rel-
evant. Clearly, many “balancing acts” are required if we are to do proper
justice to all the data which is potentially relevant to a cognitive linguist, so
understood. Cognitive semantics is a subfield of Cognitive Linguistics, but
is hardly less general, concerned as it is with “investigating the relationship
between experience, the conceptual system and the semantic structure en-
coded by language” (Evans and Green 2006: 48). The three phenomena
referred to in this characterization of Cognitive Semantics – experience, the
conceptual system, semantic structure encoded by language – also suggest a
80 John Newman

variety of techniques and methodologies for their full elaboration. In other

words, whether one is pursuing Cognitive Linguistics or a more narrowly
defined Cognitive Semantics, there will have to be some balancing acts
required on the part of the researcher in order to reconcile the diversity of
evidence.
Rather than attempt an overview of all the issues we face as we try to
navigate our way through all the possible types of data and methodologies
which Cognitive Linguists or Cognitive Semanticists might avail them-
selves of, I will focus on one kind of data and the methodologies associated
with it, namely corpus data. Here, corpus data is understood as the collec-
tion of relatively large amounts of connected (transcribed) speech or writ-
ing. In line with the broad definition of Cognitive Linguistics above, this
kind of data and its associated methodologies are allowed for in the practice
of Cognitive Linguistics and Cognitive Semantics. Indeed, Evans and
Green (2005: 108) draw attention to the importance of usage-based data in
Cognitive Linguistics, a point taken up further in Section 2 below: “lan-
guage structure cannot be studied without taking into account the nature of
language use”. Geeraerts (2006: 29) speaks of the “growing tendency of
Cognitive Linguistics to stress its essential nature as a usage-based lin-
guistics”. In light of this particular claim about Cognitive Linguistics, it
follows that corpus data drawn from actual usage is not only permissible,
but necessary in cognitive linguistic and Cognitive Semantic research.1
This is a point worth clarifying at the outset, since one could imagine that a
field called “Cognitive Linguistics” or “Cognitive Semantics” would not
possibly be concerned with the external products of usage. Teubert (2005:
8), for example, seems to understand “Cognitive Linguistics” in this way
and, not surprisingly, considers corpus linguistics and Cognitive Linguistics
to be “two complementary, but ultimately irreconcilable paradigms”. The
reality of practice in the field known as Cognitive Linguistics, however, is,
for better or for worse, the reality captured by the characterization quoted
above from Evans and Green (2006), whereby every attempt is made to
explore and reconcile a range of evidence from quite varied sources (cf.
Tummers, Heylen, and Geeraerts 2005).
Evans and Green (2006: 153-467) provides a full overview of the issues
which are of interest to Cognitive Semanticists. It would be impossible to
review the use of corpus techniques as they might apply to all these issues.
Instead, I will address a number of quite fundamental issues which are rel-
evant to the use of corpora in linguistics generally, but particularly with
respect to Cognitive Linguistics. In what follows, I question a number of
 Empirical pursuits in Cognitive Linguistics 81

current issues: over-reliance on language data abstracted from the context

of communication (Section 2), methodological issues involving registers
(Section 3), inflected forms vs. lemmas (Section 4), reliance on part of
speech tags (Section 5), and use of statistical measures (Section 6). My
intention is to draw attention to alternative ways of carrying out corpus-
based studies, alternatives which linguists of a cognitive linguistic persua-
sion, as understood by Evans and Green, should be open to. While the
focus of this volume is Cognitive Semantics, I will address issues which, I
believe, have relevance for the larger field of Cognitive Linguistics, not just
for Cognitive Semantics.

2. The communicative context

One issue which deserves more discussion than it receives in Cognitive

Linguistics, and more generally linguistics, is the scope of our inquiry.
Despite all the subfields of linguistics that exist, the field as a whole re-
mains preoccupied with the study of abstract structure at the expense of the
study of communication. Indeed, the study of abstract linguistic structure
has come to virtually define the discipline of linguistics. This imbalance
will arguably always remain a feature of the field, but it does not follow
that the study of communicative activity should be ignored to quite the
extent that it is. Cognitive linguists have been more open to incorporating
contextual information in their analyses than linguists from most other
schools of thought. Even with this openness to a more inclusive view about
language and context, however, the study of the entire communicative con-
tet seems to elude cognitive linguists.
The “usage-based hypothesis” already alluded to is sometimes held to
be a key idea guiding research in Cognitive Linguistics and one might ex-
pect that hypothesis to lead to the study of communication at a primary
concern. Evans and Green (2006) elevate the usage-based hypothesis to one
of two fundamental assumptions in Cognitive Linguistics:
(1) a. The first guiding assumption [of the cognitive approach to grammar] is
the symbolic thesis which holds that the fundamental unit of grammar is a
form-meaning pairing or symbolic unit. (p. 476)
b. The second fundamental assumption of the cognitive approach to
grammar is the usage-based thesis... [this thesis] holds that the mental
grammar of the speaker ...is formed by the abstraction of symbolic units
from situated instances of languages use. (p. 478)
82 John Newman

Evans and Green (2006) point out that their usage-based hypothesis leads to
a rejection of any strict division between “competence” and “performance”,
or knowledge of language and use of language. While this view would cer-
tainly allow for a broadening of the scope of inquiry, usage as such has not
been a particularly common research focus in the field as a whole. In some
other publications which purport to give an overview of Cognitive Lin-
guistics, usage does not occupy any privileged place. Even in its second
edition, Dirven and Verspoor’s (2004) introduction to the field has no ex-
tended discussion of the concept and the term usage is not listed in the in-
dex. The closest they come to a discussion of usage is a review of Gricean
maxims and Searle’s notions of speech acts which, though relevant to the
study of communicative context, hardly constitute any dramatic shift of
attention in the scope of inquiry.
If we are going to allow a greater role for the communicative context in
our study of language, then we need to be open to a larger range of data
than we are accustomed to. The language corpora we are in the habit of
using do not normally encompass contextual information, e.g., intonation,
pauses, eye movement, etc. in the case of conversational exchanges. It is, of
course, possible in theory to include such information in a corpus (see, for
example, McEnery, Xiao, and Tono 2006: 40-41 for a discussion of models
of pragmatic annotation), but such corpora are not widely used. The pre-
dominant interest remains in the textual or verbal component of communi-
cation, abstracted from context, rather than the communicative activity
which is taking place in the context. Hunston (2002: 94) is fairly typical of
most linguists (cognitive or otherwise) in terms of her priorities: “However
much annotation is added to a text, it is important for the researcher to be
able to see the plain text, uncluttered by annotation labels. The basic pat-
terning of the words alone must be observable at all times [in the corpus
data].” [my italics]
If cognitive linguists are to pursue the study of “situated instances of
language use” (cf. Greens and Evans above), then one might reasonably
hope for a broadening of the field of inquiry (cf. Wichmann 2007: 82-83 in
which the author calls for data from all channels of communication to be
included in our corpora). Charles Goodwin in publications such as Good-
win (1979, 1980, 1981) offers a glimpse into what it might mean to study
situated instances of language use. He goes much further than we are accus-
tomed to in traditional “conversational analysis”, directing attention to the
psychological and social processes which occur during conversation: the
direction of gaze on the part of participants and when one participant’s gaze
 Empirical pursuits in Cognitive Linguistics 83

meets another participant’s eyes, withdrawing a gaze, the lowering of the

volume of the voice, whether one can be certain that a stretch talk was
heard by other participants, etc. This kind of research program represents
the study of situated instances of language par excellence and cognitive
linguists have much to learn from that research.2

3. Registers

We can be grateful that, along with increased use of corpora in linguistics,

there has been an increased awareness of the relevance of registers, or
genres, to the description of linguistic phenomena. One does not need to
look further than such well-known corpora as the British National Corpus
(BNC), the American National Corpus (ANC), and the International Corpus
of English (ICE) to appreciate how widespread the differentiation of genres
has become. Within the field of Cognitive Linguistics, there is, I believe,
widespread recognition of the need to separate out genre-specific behav-
iours, a practice which needs to be maintained if we are concerned with
probing more deeply into fine-grained semantic properties of words or con-
structions. Sometimes, there can be valid reasons for working with an entire
corpus such as ICE without distinguishing register. Searching the whole of
ICE would be justified, for example, if one is simply exploring all possible
contexts of a search term, or all possible nuances carried by a search term,
especially at an initial, exploratory stage of an investigation.
A “spoken vs. written” distinction has become so commonplace in cor-
pus work that it seems unremarkable as a major division. However, it is by
no means a simple separation and the dangers of too easy an acceptance of
such a dichotomy are worth repeating here, even if they are obvious. There
is spoken language which derives from fully premeditated, edited, revised,
and polished written language designed to be read, e.g., formal speeches
delivered to an audience, formal lectures, news broadcasts, etc. The lan-
guage of such performances has much more in common with highly formal
writing which might appear in the written part of a corpus than it does with
more spontaneous forms of oral communication. Conversely, there are very
spontaneous forms of written language which share much, stylistically,
with conversation, e.g., internet blogs, internet chat room talk, the styles of
some personal diaries. Spontaneous conversation would seem to occupy a
special place among all the genres in so far as it is represents a relatively
basic kind of human interaction.
84 John Newman

One does not necessarily have to agree that face-to-face conversation is

paramount in terms of our communicative activities (and it may not be for
some individuals) to accept that it is an important kind of human activity
and deserving of study, especially for cognitive linguists. Arguably, it is
spoken, face-to-face conversation which is the genre which needs most to
be compared and contrasted with the written genres. It is all the more
frustrating, then, that spontaneous, face-to-face conversation remains so
underrepresented as a genre in electronic corpora, even in “spoken” cor-
pora. The BNC contains approximately 4.2 million words of such conversa-
tion and, as such, remains one of the largest resources for this genre of Brit-
ish English. There are many reasons for this: ethical issues surrounding
speakers, difficulty of understanding the exact nature of conversation with-
out access to paralinguistic features, difficulty of transcribing even the tex-
tual part of conversation etc. Some spoken corpora may contain little or no
truly spontaneous conversation and offer little or no insight into the forces
at work in ordinary face-to-face conversation.
Hand in hand with the relative rarity of conversational corpora, many of
the tools that have been designed to help linguists work better with elec-
tronic corpora do a poor job of searching and retrieving in conversational
transcripts. Searching for forms and their contexts in the utterances of a
particular speaker is not impossible but remains a complicated business in
commonly used corpus tools such as Wordsmith Tools
(www.lexically.net/wordsmith/, Scott 2004). Retrieving forms and larger
contexts including the previous utterance/turn and the following utter-
ance/turn is even more difficult in terms of built-in routines in commonly
used software. The CLAN tool used with transcriptions following the
CHAT format is probably still the most effective tool for this kind of search
and retrieval (and free), even though it has been twenty years since these
resources first became available. ELAN (http://www.lat-mpi.eu/tools/elan)
also offers convenient separation of tiers for speakers and integration with
audio/video, making it very suitable for the study of conversation. The
North Carolina Sociolinguistic Archive and Analysis Project, SLAAP
(http://ncslaap.lib.ncsu.edu/index.php), provides a unique model for the
integration of audio and text in an online environment (for an overview, see
Kendall 2007). As well as allowing the user to listen to or download an
audio file (in this case, interviews), SLAAP has integrated audio and tran-
script so that the user can undertake an acoustic analysis of any line of the
transcript (corresponding to an unbroken stretch of speech). The acoustic
analysis is made possible by Praat scripts which run “on the fly” and no
 Empirical pursuits in Cognitive Linguistics 85

additional software is required on the user’s computer. SLAAP offers users

a welcome choice of formats for the transcript: vertical format with each
successive speaker beginning a new row; column format in which utter-
ances by each speaker are shown in one column; a Henderson Graph format
in which the duration of each utterance corresponds to the length of a hori-
zontal line, together with the beginning and end times of the utterance (for
more details, see Kendall 2007 and Newman 2009).
There are other ways in which tools may build upon, and be constrained
by, a “spoken” vs. “written” dichotomy, potentially leading the researcher,
in subtle and unintended ways, away from an interest in face-to-face con-
versation. BNCWeb (http://www.bncweb.info/) is a case in point. While
BNCWeb is an extremely attractive interface for working with the BNC,
and one that I would strongly recommend, it nevertheless has certain inbuilt
biases which might encourage a researcher to explore (all) spoken vs. (all)
written contrasts, rather than, say, face-to-face conversation vs. written
genres. The built-in features I have in mind are: (1) BNCWeb includes pre-
compiled frequency lists for spoken vs. written, not for face-to-face conver-
sation; (2) word association measures (MI, z-score etc.) can be obtained
reliably for spoken vs. written, not for face-to-face conversation; (3) pro-
cessing speed is best for all the corpus, spoken sub-corpus, or written sub-
corpus. To be fair, one should point out that the BNC itself attaches most
importance to the spoken vs. written distinction and, in a sense, BNCWeb
is just allowing the user to access the corpus following the inherent major
genre division in the BNC. And one must also add that the BNCWeb does
allow a user to investigate numerous features of face-to-face conversation
as a genre in its own right.

4. Tags

POS tags and syntactic relation markup are basic to some corpus linguistic
methods, e.g., Collostructional Analysis. But neither the parts of speech nor
syntactic relations are uncontroversial (neither for English, nor for other
languages) and analyses which appeal uncritically to such concepts invite
criticism.
The part of speech categories available for English depend upon tradi-
tions of analysis and classification, and there are different traditions (differ-
ent degrees of abstraction, different terms, different degrees of sub-
classification, relevance of semantic vs. syntactic criteria in the classifica-
86 John Newman

tion etc.). Arguably, the most important role for linguists should be the
critiquing and refinement of these classifications into parts of speech for
English and other languages, rather than the acceptance of any one of these
systems of parts of speech. Huddleston’s (1980) detailed summary of the
different structural properties of thirty English auxiliary and modal verbs is
still instructive in the manner in which the author demonstrates how each of
these verbs behaves in its own unique way with respect to a set of structural
properties. Huddleston considers a total of thirty-seven properties which
include a number of properties less commonly appealed to in the literature,
but which are no less interesting or important because of that, e.g., whether
the verb can carry an emphatic stress to make a polarity contrast, whether
the verb can occur immediately before a ‘deletion’ site, and whether the
verb can take a negative complement. Huddleston’s summary of properties
makes it clear that there are many subtle ways in which this set of verbs
may be subcategorized. Indeed no two verbs display exactly the same prop-
erties. His discussion is a reminder of the important role that linguists have
to play in directing attention to the complexities and obstacles associated
with part-of-speech classifications. Cognitive linguists should be prepared
to problematize part of speech systems rather than simply accept them as
given. Similar observations hold for the related issue of grammatical rela-
tions within a clause and how these are determined. Croft (2001) strikes me
as a good example of the kind of original research that cognitive linguists
could be carrying out on both parts of speech systems and grammatical
relations, in a manner very compatible with the tenets of Cognitive Lin-
guistics. This is not to say that linguists should never appeal to the tradi-
tional categories, only that they should be constantly aware of alternatives
and hence cautious about claims based on such categories.
Quite apart from the issue of deciding upon a particular system for parts
of speech or grammatical relations, there is the issue of how successfully a
tagging algorithm captures the relevant facts. Consider the tags assigned to
the singing forms in the examples in (2) and (3), taken from the BNC.
(2) a. Cos I kept singing it this morning. VVG
b. She says she couldn’t stop singing. VVG-NN1
(3) a. The singing nearly raised the roof. NN1-VVG
b. Sandy can lead the singing. NN1
VVG refers to the –ing form of a lexical verb, NN1 refers to a singular
noun, and a hyphenated tag indicates an ambiguous result with the more
likely tag appearing first. In (2a), it is presumably the presence of the direct
 Empirical pursuits in Cognitive Linguistics 87

object it immediately after singing that contributes to an unambiguous

VVG tag assignment, whereas the absence of any direct object phrase in
(2b) leads to the ambiguous tag assignment. It is not clear just what one
should make of this difference in tagging. Accept it as a potentially import-
ant difference or disambiguate it in favour of one of VVG or NN1? In any
case, it is up to the researcher to make the decision; the tagging of the cor-
pus does not in any way “solve” the linguistic problem of determining parts
of speech. The tag assignment of NN1-VVG in (3a) is presumably related
to the occurrence of the adverb nearly immediately after singing. Nearly
modifies the following verb raised, not singing, in (3a) and should not have
any effect on how singing is analyzed for its part of speech. A linguist
would surely analyze singing in an identical way in (3a) and (3b). Again,
the tags of a corpus are only a starting point for analysis, not the end point.
The experience of working with markup of grammatical relations re-
ported in Gries, Hampe, and Schönefeld (2010) highlights the dangers of
relying, uncritically, on markup within the corpus. The authors were inter-
ested in properties of an “as-predicative” in ICE-GB, exemplified in (4):
(4) a. I do not regard the Delors Report as [NP some kind of sacred text].
b. Such a force could never be described as [AdjP purely deterrent].
c. We see the hard ECU as [S being extremely useful].
Grammatical relations can be explored in ICE-GB using the accompanying
ICECUP program distributed with the corpus. The authors extracted all
occurrences of the relevant search pattern [VP Vcomplex transitive [PP as]]
using ICECUP. After some manual correction of the output, the authors
found a total of 687 tokens of the as-predicative construction. As the
authors point out, though, this number is misleading, as the search method
fails to identify many more instances of this construction in the corpus. By
searching for [PP as], followed by manual identification of all occurrences
of the as-predicative, the authors retrieved 1,131 tokens of the construction,
a full 65% more than in the first attempt relying upon the “complex transi-
tive” parse in the corpus.

5. Lemma vs. inflected forms

It has been become commonplace, at least in English corpus linguistic re-

search, to report on the behavior of lemmas (e.g. GET), as opposed to the
inflected forms that constitute the lemma (e.g., get as an infinitive, get as a
88 John Newman

present tense form agreeing with a plural subject, gets as a present tense
form agreeing with a singular subject, got as a past tense form etc.). Recent
studies, however, have revealed interesting patterning around particular
inflected forms, as opposed to lemmas, suggesting that investigating lan-
guage at the inflectional level is a promising line of inquiry.3 In a study of
REMEMBER in spoken corpora, Tao (2001, 2003) discusses the prominence
of the simple present tense forms of this verb, used with a first person sin-
gular subject (I remember) or a null subject (remember) and suggests a
grammaticalization process is under way, confined to particular inflected
forms of the verb. Scheibman (2001), in a study of informal conversation,
found that 1st singular and 2nd singular subjects occur with particular verbs
of cognition with a relative high frequency (I guess, I don’t know, you
know, I mean) reflecting a particular pragmatic value for such combinations
in conversation. Scheibman (2001: 84) also emphasizes the need to exam-
ine “local” patterns in grammatical research and cautions against relying
just on the superordinate grammatical categories (person, verb type, tense
etc.). Deignan (2005: 158-159) discusses the different evaluations or “pro-
sodies” that attach to singular rock vs. plural rocks in the sub-corpora of the
Bank of English, reporting a tendency for the singular to have positive ev-
aluations (as in the rock on which our society is built) while the plural tends
to have negative evaluations (their marriage has been on the rocks for a
while). Newman and Rice (2004) report on how the inflectional differences
between the –ing and past tense forms in the pairs sitting and.../sat and...,
standing and.../stood and…, lying and.../lay and... influence the range of
following verbal collocates. Newman and Rice (2006) explore the different
constructional patterns found with positive, comparative, and superlative
forms of adjectives.4 Table 1 summarizes the key results from that study for
the pair slight and slightest. The table lists the top 25 noun collocates in
Adjective + Noun combinations for these two forms and the frequency of
occurrence of the collocation. The results are sorted according to log-
likelihood, as calculated by BNCWeb. The two lists of nouns show differ-
ent tendencies in their semantic fields. With the positive form slight, the
majority of nouns in the list have meanings related to ‘change, variation’,
e.g., increase, variation, modifications etc. With the superlative form
slightest, on the other hand, the majority of nouns have meanings from the
domain of ‘cognition, perception, intention’, e.g., hint, idea, doubt etc. The
different semantic tendencies of the noun collocates in these lists can only
be appreciated when the researcher is investigating words at the level of the
inflected forms.
 Empirical pursuits in Cognitive Linguistics 89

Table 1. Top 25 noun collocates of slight and slightest in the whole BNC, with
frequencies of collocation, sorted by log-likelihood (LL) value, as calcu-
lated by BNCWeb. Words in bold show the preference for ‘change’
nouns with slight and ‘cognition, awareness’ nouns with slightest.
slight N Freq. LL slightest N Freq. LL
increase 57 499 hint 25 338
smile 48 455 bit 36 303
variations 28 284 movement 29 271
angle 27 274 idea 25 203
improvement 28 256 doubt 22 198
modifications 20 234 sign 20 192
acidity 17 231 chance 21 184
breeze 20 215 interest 24 181
change 34 213 intention 15 152
delay 20 190 provocation 10 145
frown 15 186 suspicion 12 142
difference 26 183 difference 14 115
shrug 14 181 thing 17 109
changes 29 181 inclination 8 106
bow 16 168 touch 11 93
decrease 15 167 breeze 8 91
exaggeration 13 165 flicker 6 79
figure 26 162 inkling 5 75
pause 16 160 trace 7 71
differences 20 145 notice 9 70
fall 18 142 degree 9 68
movement 22 140 sound 9 62
rise 19 136 nod 5 58
variation 15 131 suggestion 6 55
flush 10 120 encouragement 5 51

Multivariate analysis techniques which incorporate the full range of related

forms of a lemma and a large number of variables associated with these
forms are promising techniques for exploring, at least initially, sub-patterns
which might well escape detection if we were to rely solely on linguistic
intuition to guide us. The clustering analysis underlying “behavioral pro-
files”, as illustrated in Gries (2006), Gries and Divjak (2009), and Gries and
Otani (forthc.) offer opportunities for exploring differential patterning be-
tween inflected forms. Gries and Otani (forthc.), for example, reveals in-
triguing clustering of positive, comparative, and superlative forms of adjec-
90 John Newman

tives. Similarly, correspondence analysis, as exemplified in Glynn (2009,

2010), enables a researcher to discern subtle patterning around related word
froms involving multiple factors.
Some corpus linguistic tools, for example BNCWeb and Sketch Engine
(Kilgarriff and Tugwell 2001; Kilgarriff, Rychly, Smrz, and Tugwell 2004),
include ways to examine the frequency and contextual information about
either inflected forms or lemma. Sketch Engine has a particularly attractive
feature of signaling when a particular inflected form has a “salient” distri-
bution. So, for example, the -ing (VVG) form waiting has a relatively high
frequency of occurrence within the set of inflected forms of the verb lemma
WAIT. Waiting occurs 8,053 times in the BNC, which is 40.62% of all
forms of the lemma WAIT which occurs 19,824 times. This puts WAIT in the
top 10% of verbs ranked by descending relative frequency of their –ing
forms (as calculated by Sketch Engine), hence waiting is a salient inflected
form of WAIT. This key result, that waiting occupies a special place among
the inflected forms of WAIT, is indicated along with other results from a
Word Sketch query without the need for any calculations on the part of the
user. Or, to take a noun example, WOMAN, the plural form women occurs
34,223 times and accounts for 60.87% of tokens of the lemma WOMAN,
putting WOMAN in the top 10% of nouns in terms of relative frequency of
their plural forms, hence this fact is indicated in the Word Sketch. The
same is not true, for example, of the plural form men which accounts for
only 37.55% of the forms of MAN. While acknowledging this attractive
feature of Sketch Engine, it should be mentioned that the feature is imple-
mented in a fairly limited way, at this point in time. Only the –ing form and
passive use is reported on in this way for verbs, and only the use of the
plural form is reported on for nouns. Furthermore, the “Sketch Differences”
tool of Sketch Engine, which allows easy comparison of constructional
possibilities of two words based on the BNC, is only implemented with
respect to differences between two lemmas–one couldn’t compare con-
structional possibilities of, say, remember and remembered, which as Tao
has shown is a fruitful area of investigation.

6. Quantitative issues

Let us start with the the question of establishing how many words there are
in a corpus – something that might appear to some to be too trivial for seri-
ous discussion. With so many sophisticated statistical techniques now
 Empirical pursuits in Cognitive Linguistics 91

available for measuring the frequencies and associations within a corpus,

one is more used to discussion of much more complex quantitative issues
corpus linguistics. However, almost all statistical calculations used in cor-
pus linguistics, from the relatively simple to the most sophisticated, in-
corporate word counts into their formulae and so establishing the size of a
corpus is not inconsequential. Nor is it a trivial task.
Establishing the number of “words” in the BNC is not straightforward.
The BNC World Edition is said to contain 100,467,090 orthographic words.
One could use this figure as a basis for quantifying the size of the corpus,
though the approximation of 100 million words is also used on occasion.
But one might also consider using the “w-units” as a basis. “w-units” are
the items tagged for parts of speech and may be smaller or larger than what
we know as orthographic words in English, as exemplified by the tagging
of out of and gonna in (5).
(5) a. <w type="PRP" lemma="out of">out of </w>
b. <w type="VVG" lemma="gon">gon</w><w type="TO0"
lemma="na">na </w>
In (1a), the sequence out of is analyzed as a single (compound) preposition,
hence one w-unit. The number of compound prepositions and compound
conjunctions which are tagged as single w-units is substantial, including
forms such as from time to time (1,573 tokens), in addition (4,229), in front
of (5,558), in general (3,847), in order (12,025), in terms of (9,238), more
than (13,149), out of (41,845), of course (24,091), rather than (19,332).5
Conversely, in (2b), the single word gonna is analyzed as two w-units the –
ing form of a lexical verb GON followed by the infinitival TO. As the ana-
lyzed units, assigned a part of speech, the w-units are a perfectly reasonable
basis for quantifying the size of the corpus, but this would be different from
the number of orthographic words.6 BNCWeb, in its calculations of statist-
ics, relies on the total number of tagged items in the corpus which includes
punctuation marks (Sebastian Hoffman, personal communication). Again,
this is not an unreasonable way to proceed–after all, punctuation such as
full stops, commas etc. are often seen as providing important contextual
clues in understanding the behavior of words in a corpus. Punctuation
marks contribute a kind of meaning to a corpus, in this view, and so de-
serve to be counted. It will be obvious that different decisions about what
unit to count affects any further calculations based on total number of units.
And, as noted in Footnote 6, even if one decides to count w-units, one is
faced with different figures even just consulting the BNC homepage.
92 John Newman

A different kind of question arises when carrying out Collostructional

Analysis. In this kind of analysis the researcher arrives at a measure of the
statistical significance of the occurrence of a lexeme L in a construction C
(cf. Stefanowitsch and Gries 2003: 218). One of the numbers required for
this analysis is the frequency of “all other constructions with lexemes other
than L” in the corpus. The question arises as to how to count the number of
relevant constructions in the corpus, a question considered also by Schmid
(this vol.). In the case of the construction [N waiting to happen], Ste-
fanowitsch and Gries choose to count the total number of verb tags as the
number of constructions with lexemes other than L. In the absence of more
detailed discussion on this point by the authors, there seem to be a number
of alternative ways to proceed when determining the number of construc-
tions to take into account, as pointed out by Schmid. One might prefer to
consider –ing forms of lexical verbs, for example, given the fact it is spe-
cifically the –ing form of WAIT which features in this construction. Or one
might prefer to consider all instances of verbs concatenated with the infini-
tive to, counting each pair of concatenated verbs as one instance of the
construction.
It is inevitable that the quantitative techniques that we use in carrying
our corpus-based research will become increasingly sophisticated and a
familiarity with quantitative and statistical techniques of analysis will sim-
ply be unavoidable. I see myself as an end-user, rather than a creator, of
statistical techniques and so for me the challenge that presents itself is one
of understanding the strengths and weaknesses of the techniques and choos-
ing between competing techniques. I have not found this particularly easy,
in part because even those who are more statistically sophisticated than I
am seem to have more questions than answers when it comes to deciding
what “best practice” is supposed to be. A case in point is Kilgarriff (2005),
published in the first volume of the journal Corpus Linguistics and Lin-
guistic Theory, and the commentary on it by one of the editors of the jour-
nal, Gries (2005). The two papers address the issue of “null hypothesis
testing” in corpus linguistics, and Gries takes up the challenge posed by
Kilgarriff’s claim that null-hypothesis significance testing (which assumes
“randomness”) leads to unhelpful or misleading results. Gries (2005: 281)
suggests ways to deal with some of the problems raised by Kilgarriff, but
concludes with the words: “On the basis of these results, it is very difficult
to decide what the right quantitative approach in such word-frequency stud-
ies may be. What is not so difficult is to notice that much more exploration
in these areas is necessary and that the results show how much we may
 Empirical pursuits in Cognitive Linguistics 93

benefit from taking up methodological proposals from other disciplines to

refine, and add to, our methods…”. The Gries commentary on the Kilgar-
riff article typifies for me the conundrum we are faced with when we come
to apply more sophisticated statistical measures, most of which assume
some kind of randomness in the population (= words in the corpus) being
measured. The application of statistical measures in corpus linguistics is
clearly still very much in its infancy and, while there are various statistical
practices that have become commonplace in corpus linguistics, one cannot
assume they represent “best practice” and one must be at least a little cir-
cumspect about all such results.
In light of the foregoing remarks, I think we should still concede a place
for the “less sophisticated” quantitative measures in our work with corpora.
As an example of a relatively unsophisticated approach, one might cite
Schmid’s measures of attraction and reliance measures which I find both
revealing and useful, even if Schmid himself describes them as “simple
arithmetic” (Schmid this volume, section 5.1). These measures, as they
apply to nouns are given below:
(6) a. Attraction = frequency of a noun in a pattern x 100
total frequency of the pattern
b. Reliance = frequency of a noun in a pattern x 100
total frequency of the noun in the corpus
Attraction measures the extent to which a particular pattern attracts a noun,
while reliance measures the extent to which noun appears in one particular
pattern versus other patterns. Schmid (this vol.) provides a helpful discus-
sion of the strengths and weaknesses of these formulae and an insightful
comparison of these measures with the results of Collostructional Analysis.
Given the controversies surrounding the use of statistical measures in cor-
pus linguistics, it seems perfectly reasonable to continue to make use of a
range of more qualitative techniques to understand the behavior of words in
context. Hunston (2002) helpfully describes a variety of useful methods for
examining usage of words in a corpus, working from concordance lines.
While some of the methods that she describes rely upon statistical measures
such as MI, z-score, and t-score, others do not. The non-statistical ap-
proaches that she discusses are still properly called “methods” which have
their own rationale and rigour. And where the researcher is exploring subtle
semantic nuances of a word as evidenced in the context, including the lar-
ger discourse context, then one has little choice but to work with these
“softer” quantitative methods. The chapters in Deignan (2005) and Ste-
94 John Newman

fanowitsch and Gries (2006), for example, contain many insightful and
thought-provoking analyses of corpus data employing little more than per-
centages and Chi-square measures.

7. Conclusion

If we accept that usage is to play a key role in cognitive linguistic research

(as claimed in Evans and Green 2006), then corpora and the analytical
techniques associated with them will necessarily be an ever more important
focus for researchers in this area. I find myself in complete agreement with
Geeraerts (2006: 45), then, when he writes that “a program in Cognitive
Linguistics worthy of that name should resolutely opt for advanced training
in empirical linguistics”, where advanced training in empirical linguistics
necessarily includes advanced training in statistical techniques.
There is an abundance of corpora available to researchers, and so one
can look forward to much more corpus-based research into Cognitive Lin-
guistics. Corpus data seem very accessible and “ready to use”, especially
with some of the corpus tools now available to researchers. But it is im-
portant to appreciate alternative ways of proceeding with corpus-based
research. In the discussion above I have emphasized the need to bear these
alternatives in mind in working with a corpus. I believe that the alternatives
I have reviewed in the discussion above are all worth pursuing further:
working with language as structure and language as communicative ac-
tivity; the inclusion of conversational language and other genres in studies
of language; investigating both inflected forms and lemmas; a tolerance for
both quantitative and qualitative methods. Our goal as cognitive linguists
should be one of maintaining a balance between alternative kinds of evi-
dence and alternative research methodologies.

Notes

1. The increasing interest in usage-based data is true of the field of linguistics

and is by no means restricted to Cognitive Linguistics. An Editor of Language
has observed, with reference to the contents of the journal: “…we seem to be
witnessing…a shift in the way some linguists find and utilize data–many pa-
 Empirical pursuits in Cognitive Linguistics 95

pers now use corpora as their primary data, and many use internet data. (Jo-
seph 2004: 382).
2. One should mention, too, the contributions of Harris (1996, 1998) who argues
forcefully for an agenda for the study of language which situates language
well and truly in the context of communication, what Harris calls an “integra-
tionist approach”. Thorne and Lantolf (2006) is a more recent statement of a
similar kind in which the authors argue for a “Linguistics of Communicative
Activity”, the goal of which is to “disinvent language understood as an object
and to reinvent language as activity...” (Thorne and Lantolf 2006: 171, italics
original).
3. Stubbs (2001: 99) draws attention, in passing, to the issue of investigating
different inflected forms, as opposed to lemmas, and cites a remark by Sin-
clair (1991: 8) suggesting the value of investigating the different forms of a
lemma in terms of their uses. Further discussion of the role of inflected forms
vs. lemmas can be found in Newman (2008).
4. Claridge (2007), investigating the superlative of English adjectives in conver-
sational English, demonstrates the value of considering specific inflected
forms of adjectives, as opposed to lemmas. She reports on a range of proper-
ties of the superlative which could not be claimed of the class of adjectives in
general.
5. The tagging system can lead to difficulties when it comes to retrieving some
of these forms. Mark Davies comments on this in connection with searches for
multiword units like in charge of when using his Contemporary Corpus of
American English (COCA) site: “…the multi-word units work only with quer-
ies that involve at least one specified word in the query. In other words,
[* charge *] will find in charge of, but a query composed of just wildcards
[* * *] or parts of speech ( [prp] [nn1] [pr*] ) would not” (http://corpus.
byu.edu/bnc/). Davies’ comment is a reminder of just how careful one must be
in retrieving information from the BNC. The user needs to be well aware of
the full range of multiword w-units in performing searches.
6. On the relevant web page of the BNC, the number of w-units in the whole
corpus is said to be “slightly less than” 97,619,934. On the same page, one
paragraph after this number is given, we are told that the total number of w-
units is 89.30 (written) + 10.58 (spoken) = 99.88 million, a discrepancy of
more than 2 million words (www.natcorp.ox.ac.uk/corpus/index.xml.ID=
numbers, accessed 3 July 2007). The BNCWeb, in its reporting on searches in
the whole corpus, gives the number 97,626,093, which presumably refers to
w-units, rather than orthographic words, but still differs from the two numbers
given on the BNC homepage.
96 John Newman

References

Chomsky, Noam
1965 Aspects of the Theory of Syntax. Cambridge, Mass.: M.I.T. Press.
Claridge, Claudia
2007 The superlative in spoken English. In Corpus Linguistics 25 Years
On, Roberta Facchinetti (ed.), 121-148. Amsterdam / New York:
Rodopi.
Croft, William
2001 Radical Construction Grammar. Oxford: Oxford University Press.
Deignan, Alice
2005 Metaphor and Corpus Linguistics. Amsterdam / Philadelphia: John
Benjamins.
Dirven, René and Marjolijn Verspoor
2004 Cognitive Exploration of Language and Linguistics, 2nd ed.
Amsterdam / Philadelphia: John Benjamins.
Evans, Vyvyan and Melanie Green
2006 Cognitive Linguistics: An introduction. Edinburgh: Edinburgh Uni-
versity Press.
Geeraerts, Dirk
2006 Methodology in Cognitive Linguistics. In Cognitive Linguistics:
Current applications and future perspectives, Gitte Kristiansen,
Michel Achard, René Dirven and Francisco José Ruiz de Mendoza
Ibáñez (eds.), 21-49. Berlin / New York: Mouton de Gruyter.
Glynn, Dylan
2009 Polysemy, syntax, and variation: A usage-based method for Cogni-
tive Semantics. In New Directions in Cognitive Linguistics, Vyvyan
Evans and Stéphanie Pourcel (eds.), 77-104. Amsterdam / Philadel-
phia: Benjamins.
2010 Synonymy, lexical fields, and grammatical constructions. A study in
usage-based Cognitive Semantics. In Cognitive Foundations of Lin-
guistic Usage-Patterns, H.-J. Schmid and S. Handl (eds.), 89-118.
Berlin / New York: Mouton de Gruyter.
Goodwin, Charles
1979 The interactive construction of a sentence in natural conversation. In
Everyday Language: Studies in ethnomethodology, G. Psathas (ed.),
97-121. New York: Irvington.
1980 Restarts, pauses, and the achievement of mutual gaze at turn-
beginning. Sociological Inquiry (Special ed. Language and Social
Interaction 50:3/4), Don Zimmerman and Candace West (eds.), 272-
302. London: Lawrence Erlbaum.
 Empirical pursuits in Cognitive Linguistics 97

1981 Conversational Organization: Interaction between speakers and

hearers. New York: Academic Press.
Gries, Stefan Th.
2005 Null-hypothesis significance testing of word frequencies: A follow-
up on Kilgarriff. Corpus Linguistics and Linguistic Theory 1(2):
277-294.
2006 Corpus-based methods and Cognitive Semantics: The many mean-
ings of to run. In Corpora in Cognitive Linguistics. Corpus-based
approaches to syntax and lexis, Stefan Th. Gries and Anatol Ste-
fanowitsch (eds.), 57-99. Berlin / New York: Mouton de Gruyter.
Gries, Stefan Th. and Dagmar S. Divjak
2009 Behavioral profiles: A corpus-based approach towards Cognitive
Semantic analysis. In New Directions in Cognitive Linguistics, Vy-
vyan Evans and Stéphanie Pourcel (eds.), 57-75. Amsterdam /
Philadelphia: John Benjamins.
Gries, Stefan Th., Beate Hampe, and Doris Schönefeld
2010 Converging evidence II: More on the association of verbs and con-
structions. In Empirical and Experimental Methods in Cogni-
tive/Functional Research, Sally Rice and John Newman (eds.), 59-
72. Stanford: CSLI.
Gries, Stefan Th. and Naoki Otani
forthc. Behavioral profiles: A corpus-based perspective on synonymy and
antonymy. ICAME Journal.
Harris, Roy
1996 Language and Communication: Integrational and segregational
approaches. London: Routledge.
1998 Introduction to Integrational Linguistics. Oxford: Elsevier.
Huddleston, Rodney D.
1980 Criteria for auxiliaries and modals. In Studies in English Linguistics
for Randolph Quirk, Sidney Greenbaum, Geoffrey N. Leech and Jan
Svartvik (eds.), 65-78. London / New York: Longman.
Hunston, Susan
2002 Corpora in Applied Linguistics. Cambridge: Cambridge University
Press.
Joseph, Brian D.
2004 The editor’s department: On change in Language and change in
language. Language 80(3): 381-383.
Kendall, Tyler
2007 Enhancing sociolinguistic data collections: The North Carolina
sociolinguistic archive and analysis project. University of Pennsyl-
vania Working Papers in Linguistics 13(2): 15-26.
98 John Newman

Kilgarriff, Adam
2005 Language is never, ever, ever, random. Corpus Linguistics and Lin-
guistic Theory 1(2): 263-276.
Kilgariff, Adam, Pavel Rychly, Pavel Smrz, and David Tugwell
2004 The Sketch Engine. Proceedings of EURALEX 2004, Lorient,
France, 105-116.
Kilgarriff, Adam and David Tugwell
2001 WORD SKETCH: Extraction and display of significant collocations
for lexicography. Proceedings of the Collocations Workshop, ACL
2001, Toulouse, France, 32-38.
McEnery, Tony, Richard Xiao, and Yukio Tono
2006 Corpus-based Language Studies: An advanced resource book. Mil-
ton Park: Routledge.
Newman, John
2008 Aiming low in linguistics: Low-level generalizations in corpus-
based research. Proceedings of the 11th International Symposium on
Chinese Languages and Linguistics (IsCLL-11), May 23-25 2008,
National Chiao Tung University, Hsinchu, Taiwan. Available at
http://johnnewm.jimdo.com/downloads.php.
Newman, John and Sally Rice
2004 Patterns of usage for English SIT, STAND, and LIE: A cognitively-
inspired exploration in corpus linguistics. Cognitive Linguistics 15:
351-396.
2006 English adjectival inflection: A radical radical Construction Gram-
mar approach. The 8th Conceptual Structure, Discourse, and Lan-
guage Conference, November 2006, San Diego.
Scheibman, Joanne
2001 Local patterns of subjectivity in person and verb type in American
English conversation. In Frequency and the Emergence of Linguistic
Structure, Joan L. Bybee and Paul Hopper (eds.), 61-89. Amsterdam
/ Philadelphia: John Benjamins.
Schmid, Hans-Jörg
this vol. Does frequency in text really instantiate entrenchment in the cogni-
tive system?
Scott, Mike
2004 WordSmith Tools, Version 4. Oxford: Oxford University Press.
Sinclair, John
1991 Corpus, Concordance, Collocation. Oxford: Oxford University
Press.
Stefanowitsch, Anatol and Stefan Th. Gries (eds.)
2006 Corpus-Based Approaches to Metaphor and Metonymy. Berlin /
New York: Mouton de Gruyter.
 Empirical pursuits in Cognitive Linguistics 99

Stubbs, Michael
2001 Words and Phrases: Corpus studies of lexical semantics. Oxford:
Blackwell.
Tao, Hongyin
2001 Discovering the usual with corpora: The case of remember. In Rita
C. Simpson and John M. Swales (eds.), Corpus Linguistics in North
America: Selections from the 1999 symposium, 116-144. Ann Arbor:
University of Michigan Press.
2003 A usage-based approach to argument structure. International Jour-
nal of Corpus Linguistics 8: 75-95.
Teubert, Wolfgang
2005 My version of corpus linguistics. International Journal of Corpus
Linguistics 10(1): 1-13.
Thorne, Steven L. and James P. Lantolf
2006 A linguistics of communicative activity. In Disinventing and Recon-
stituting Languages, Sinfree Makoni and Alastair Pennycook (eds.),
170-195. Clevedon: Multilingual Matters.
Tummers, José , Kris Heylen, and Dirk Geeraerts
2005 Usage-based approaches in Cognitive Linguistics: A technical state
of the art. Corpus Linguistics and Linguistic Theory 1(2): 225-261.
Wichmann, Anne
2007 Corpora and spoken discourse. In Corpus Linguistics 25 Years On,
Roberta Facchinetti (ed.), 73-86. Amsterdam / New York: Rodopi.
Does frequency in text instantiate entrenchment in
the cognitive system?

Hans-Jörg Schmid

Abstract
This paper investigates the relation between observed discourse frequencies of
linguistic elements and structures, on the one hand, and assumptions concerning
the entrenchment of these units in the minds of speakers, on the other. While it is
usually assumed that there is a fairly direct correlation between frequency of use
and degree of entrenchment, it is argued that many essential questions concerning
this relation have remained unanswered so far: What is the role of absolute fre-
quency of occurrence as opposed to frequency relative to a given reference con-
struction? How are relative discourse frequencies to be captured statistically in
such a way that, for instance, rare lexical items that typically occur in certain con-
structions can be differentiated from frequent lexical items which are more versa-
tile but also observed to occur in the same construction, often with considerable
absolute frequencies of occurrence? What are the psychological implications of
different combinations of high and low absolute and relative frequencies? While
the paper suggests solutions to some of these problems it also points to a number of
unresolved issues to be addressed in the future and calls for a more modest and
cautious way of interpreting quantitative observations in cognitive terms.
Keywords: discourse frequency, entrenchment, quantitative approaches, collostruc-
tion

1
1. Introduction

It is common practice in corpus linguistics to assume that the frequency

distribution of tokens and types of linguistic phenomena in corpora have –
to put it as generally as possible – some kind of significance. Essentially,
more frequently occurring structures are believed to hold a more prominent
place, not only in actual discourse but also in the linguistic system, than
those occurring less often.
Cognitively-oriented corpus linguists also subscribe to this assumption,
but they tend to go one step further. Given their cognitive leaning, they
should be on the hunt for psychologically plausible models of language
102 Hans-Jörg Schmid

based on quantitative observations of corpus data. More specifically, they

try to correlate the frequency of occurrence of linguistic phenomena (as
observed in corpora) with their salience or entrenchment in the cognitive
system. A corollary of this assumption is that patterns of frequency distri-
butions of lexico-grammatical variants of linguistic units correspond to
variable degrees of entrenchment of cognitive processes or representations
associated with them.
Among early investigations pursuing this line of corpus-based cogni-
tive-linguistic reasoning are Rudzka-Ostyn’s (1989) study of the polysemy
of the English verb ask, Dirven’s (1991) paper on agree and my own work
2
on start vs. begin and the polysemy of the noun idea (Schmid 1993). Dur-
ing this early period of quantitative Cognitive Semantics, frequency pat-
terns tended to be interpreted in terms of typicality of meaning, with the
most frequent usage-patterns being taken to reflect (proto-)typical senses of
3
lexemes. Schematic meanings were also taken into account in these stud-
ies, but it seemed less clear how they are related to frequency distributions
(cf. Schmid 1993: 218). The work by Rudzka-Ostyn, Dirven and myself
was very much inspired by ideas on the network nature of (lexicalized)
conceptual categories developed in a non-quantitative framework by
Brugman (1981), Lindner (1982), Geeraerts (1983), Herskovits (1986),
Lakoff (1987), Schulze (1988) and others in the course of the 1980s.
Tracing back the historical roots of corpus-driven quantitative Cognitive
Semantics is not an end in itself or just an homage to the scholars who have
built the foundation that made the current work in the field possible. It is
important because it reminds us of the ultimate aim of cognitive linguists to
come up with linguistic models that actually claim to reflect (what we be-
lieve to know about) the way our minds work. In my perception, this aim is
in danger of falling into oblivion. Trapped in a numerical maze by the irre-
sistible lure of masses of data, smart corpus queries, long periods of num-
ber-crunching and skilful applications of advanced statistical methods,
present-day corpus-driven Cognitive Semantics runs the risk of losing sight
of the bigger picture hidden in language, and of the cognitive aspects of
meaning it purports to unravel, if only by the name by which it presents
itself to the wider linguistic community.
In view of this danger, this paper backtracks a number of steps in the
development of corpus-driven Cognitive Semantics. It will question one of
the main assumptions underlying, more or less tacitly or explicitly, much of
the current work in the field. More specifically, the paper brings under
scrutiny the hypothesis that frequency in text more or less directly instanti-
 Does frequency in text instantiate entrenchment? 103

ates salience or entrenchment in the cognitive system, a claim most force-

fully proposed in the form of my own From-corpus-to-cognition-principle
(Schmid 2000: 39) as an extension to Halliday’s dictum that “frequency in
text instantiate[s] probability in the [linguistic] system” (Halliday 1993: 3).
To make this quite clear, the aim of this contribution is not to question the
importance and significance of quantitative approaches in Cognitive Lin-
guistics as such, far from it, but rather to point out a number of potential
pitfalls and shortcomings that will have to be addressed in the future.

2. The methodology of quantitative Cognitive Semantics: An idealized

outline

To locate the issues addressed in this paper in the methodological landscape

of quantitative Cognitive Semantics, it is important to spell out the method-
ological steps typically gone through in studies set in this framework. It
should be borne in mind that this is a generalized and idealized version of
the state of the art, which may not have been applied this way in any con-
crete study but still represents the blueprint of a recipe underlying many.
The sequence of steps is summarized in Table 1 (cf. also Tummers, Heylen
and Geeraerts 2005: 238-245):

Table 1. Idealized version of the methodology of quantitative Cognitive Semantics

1. Choice of object of study: Find and define an interesting linguistic
phenomenon.
2. Choice of corpus: Find either a huge or a tailor-made special-
purpose corpus – ideally representative of the kind of language you
want to study (whatever that means) – that includes a large number
of diverse instances of the linguistic phenomenon.
3. Formulate corpus queries: Operationalize the linguistic problem in
such a way that it can be formulated as a corpus query or set of cor-
pus queries.
4. Retrieve and clean up material: Use the query to retrieve all in-
stances of the phenomenon from the corpus and get rid of clear un-
wanted hits.
5. Get material under control: Analyse the distribution and frequen-
cies of variants of the valid instances of the phenomenon retrieved;
annotate them accordingly or insert them in a database.
104 Hans-Jörg Schmid

6. Analyse material mathematically/statistically: Capture the fre-

quency distribution mathematically and check for statistical signifi-
cances.
7. Interpret quantitative findings: Interpret quantitative findings in
terms of semantic and cognitive organization.

While it is well known in corpus-linguistic circles that not a single one of

these steps is trivial, this is easily overlooked. Although a lot more could be
said about the other steps and the way they influence the outcome of corpus
studies, I will focus my attention on steps 6 and 7, since my main concern
is the significance of frequencies. Suffice it to recall at this stage that the
apparently so objective quantitative approach has a much larger number of
subjective decisions and sources of errors built into its methodological
apparatus than most practitioners of the art are willing to admit (cf.
Tognini-Bonelli 2001: 178, Mukherjee 2005: 71–72).

3. The constructions serving as case studies: Shell-content construction

The focus of this paper is on methodological issues. Since these should not
be discussed in a linguistic vacuum, I will go back to data from my own
previous research (cf. Schmid 2000) to illustrate the problems at hand. Four
related types of nominal constructions, all consisting of an abstract noun
and a complementing that-clause or to-infinitive, will serve as case studies.
The constructions are illustrated in Table 2:
Table 2. Variants of the shell-content construction
a) N + that-clause: The fact that abstract nouns are difficult to pin down ...
b) N + to-infinitive: The idea to illustrate the patterns investigated ...
c) N + BE + that-clause: The problem is that there is a lot to study.
d) N + BE + to-infinitive: The solution is to focus on a bunch of examples.
As the table shows, the four types differ with regard to the form of the
complement (that-clause vs. to-infinitive) and the link between N and com-
plement (direct link as nominal postmodifier vs. link by means of the
copula be). In my previous work I have referred to the nouns in these con-
structions as shell nouns, because they conceptually encapsulate the com-
plex pieces of information expressed in the clauses (referred to as shell
contents). The whole constructions are seen as variants of a more schematic
 Does frequency in text instantiate entrenchment? 105

shell-content construction (Schmid 2007a, Ungerer and Schmid 2006: 248-

4
250).

4. Data source and retrieval

The data re-considered here were originally collected in 1996. The material
was taken from the British section of COBUILD’s Bank of English, amount-
ing at that time to 225 million running words from the following subcor-
pora: spoken conversation, transcribed BBC recordings, ephemeral texts
such as brochures and leaflets, fiction and non-fiction books, magazines
(both lifestyle and political, including The Economist), broadsheet news-
papers (The Times, The Guardian and The Independent), the tabloid Eng-
lish Today as well as the science journal The New Scientist. It should be
borne in mind that with a proportion of about two-thirds of the whole ma-
terial, texts from media sources make up the bulk of this corpus.

Table 3. Corpus queries and numbers of matches (Schmid 2000: 44-45)

Query statement Number of matching lines
in the 225m corpus
Pattern N-cl:
NN+that/CS (NN = noun, CS = conjunction) 280,217
NN+to+VB (VB = base form of verbs) 560,148

Pattern N-be-to
NN+is+to 28,463
NN+was+to 12,728
NN+has+been+to 962
NN+will+be+to 960
NN+would+be+to 1,421
NN+would+have+been+to 133

Pattern N-be-that
NN+is+that 37,155
NN+was+that 9,104
NN+has+been+that 433
NN+will+be+that 178
NN+would+be+that 264
NN+would+have+been+that 19
106 Hans-Jörg Schmid

In the corpus queries, the nominal slots were defined by a part-of-speech

dummy (NN), while the complement clauses were identified by means of
the complementizers that and to. The queries aimed at the constructions
containing the copula included different morphological variants. Table 3
gives a summary of the query statements and the number of matching lines
in the 225 million-word corpus. In addition, analogous patterns with wh-
clauses were retrieved (e.g. the question why he didn’t come or the problem
how to define the task), as well as the highly frequent anaphoric instances
of shell nouns, but neither of these types is under consideration here.
After half-manual data-cleaning, the following numbers of valid tokens
and noun types represented in the corpus remained for further investigation.

Table 4. Valid hits in the corpus study conducted by Schmid (2000)

CONSTRUCTION Tokens Noun types

N + that-clause 141,476 350
N + to-infinitive 228,165 200
N + BE + that-clause 30,992 366
N + BE + to-infinitive 21,876 162

5. Capturing the data

Following the rationale outlined in Table 1, the next step is to arrange the
material in such a way that patterns of distribution become visible. This
step is often combined with step 6, a first attempt to capture the distribution
of the data mathematically.
Applied to the present material, the aim of steps 5 and 6 is to come up
with interesting, data-driven observations concerning the interaction be-
tween the four types of constructions and the types of nouns occurring in
them. More specifically the following questions are of concern:

– To what extent do the different types of constructions attract specific

types of nouns?
– To what extent do certain nouns depend on one or more of the con-
structions for their occurrence in actual discourse?
– Is there a semantic affinity between the frequency distribution of
types of nouns and the types of constructions?
 Does frequency in text instantiate entrenchment? 107

In what follows, two different attempts to answer these questions will be

sketched out: the attraction-reliance method proposed by Schmid (2000)
and the collostructional method introduced by Stefanowitsch and Gries
(2003).

5.1 Simple arithmetic: attraction-reliance method

In the late 1990s, when the material for the study reported in Schmid
(2000) was collected and analysed, a number of linguistico-statistical
measures, which were designed to answer questions of this type, were of
course firmly established. The most commonly applied ones were t-score
and Mutual Information (see Church and Hanks 1990), which were both
used at COBUILD (Clear 1993) and for analyses of the British National
Corpus. However, these measures were so technical that even linguists who
had applied them with some success admitted they were not able to see
behind the formulas and to interpret the actual linguistic significance (cf.
also Stubbs 1995 for a critical discussion). Moreover, these formulas are
designed to be calculated for a set of potential collocators of the so-called
node within a defined span, without taking into consideration their relation
to the node. This procedure, which may be useful and necessary if one is
concerned with co-occurrence tendencies in general, would have unneces-
sarily blurred issues that were perfectly clear in the case of shell-content
constructions.
In view of these disadvantages of the established instruments in the stat-
istical toolbox, I decided to come up with much simpler but more transpar-
ent arithmetic ways of capturing the interaction between nouns and con-
struction. The measures were called attraction and reliance and calculated
as represented in Figure 1:

frequency of a noun in a pattern x 100

Attraction =
total frequency of the pattern

frequency of a noun in a pattern x 100

Reliance =
total frequency of the noun in the corpus

Figure 1. Calculating the measures of attraction and reliance (Schmid 2000: 54)
108 Hans-Jörg Schmid

As the figure shows, attraction is calculated by dividing the frequency of

occurrence of a noun in a pattern by the frequency of the pattern in the cor-
pus. The result of this division measures the degree to which a pattern at-
tracts a particular noun. Since the denominator of the fraction is the same
for all nouns which occur in a pattern, the scores for this value are directly
proportional to the raw frequencies of nouns. The measure facilitates the
comparison of the relative importance of individual nouns for a pattern.
While capturing the relation between nouns and pattern, this is very much a
paradigmatic way of looking at the nominal slot in the pattern. Differences
in attraction are illustrated in Figure 2 for the frequency scores found for
the two nouns fact and idea in the patterns N + that-clause:
26106 x 100
Attraction fact that = 141476 = 18.45%

4812 x 100
Attraction idea that = = 3.40%
141476

Figure 2. Exemplifying differences in attraction

As shown in Figure 1, reliance is calculated by dividing the frequency of

occurrence of a noun in a pattern by its frequency of occurrence in the
whole corpus. This measure expresses the proportion of uses of nouns in
the patterns vis-à-vis other usage-types of the same noun. As the denomina-
tor of the fraction varies with the overall frequency of a noun in the corpus,
scores for reliance are not proportional to their frequency of occurrence in
the constructions. Viewed from the nouns’ perspective, reliance is a syn-
tagmatic rather than a paradigmatic measure, since it accounts for combina-
tions of nouns with types of patterns. Figure 3 illustrates differences in
reliance scores in the pattern N + that-clause for the nouns fact and realiza-
tion.
26106 x 100
Reliance fact that = = 38.13%
68472

820 x 100
Reliance realization that = 1185 = 69.20%

Figure 3. Exemplifying differences in reliance

 Does frequency in text instantiate entrenchment? 109

While fact was found much more often in the pattern N + that-clause than
realization, the latter can boast a much higher score for reliance because its
overall frequency of occurrence in the corpus is much lower, too.
The attraction-reliance method of capturing the data allows for two dif-
ferent types of information about frequency of occurrence to be made. On
the one hand, focusing on how the nominal slots in the constructions are
filled, different noun types can be ranked according to their scores for both
attraction and reliance in a certain pattern. An illustrative extract from such
a ranking for the pattern N + that-clause is given in Table 5, where the
three columns on the left-hand side provide the ranking for attraction and
the other four columns the one for reliance.
Table 5. Ranking of attraction and reliance scores for the construction N + that-
clause (top 20; 141,476 tokens in the corpus)
Noun FREQ. IN Attraction Noun FREQ. IN FREQ. IN Reliance
PATTERN PATTERN CORPUS
fact 26,106 18.45% realization 820 1,185 69.20%
evidence 5,007 3.54% proviso 111 250 44.40%
idea 4,812 3.40% assumption 1,391 3,151 44.14%
doubt 4,010 2.83% assertion 596 1,492 39.95%
belief 3,696 2.61% belief 3,696 9,344 39.55%
view 3,532 2.50% insistence 796 2,069 38.47%
hope 2,727 1.93% fact 26,106 68,472 38.13%
news 2,572 1.82% premise 274 765 35.82%
feeling 2,511 1.77% misapprehension 44 123 35.77%
impression 2,279 1.61% suggestion 2,033 5,854 34.73%
possibility 2,232 1.58% dictum 84 249 33.73%
claim 2,194 1.55% stipulation 48 145 33.10%
suggestion 2,033 1.44% misconception 91 284 32.04%
speculation 1,922 1.36% truism 47 150 31.33%
knowledge 1,794 1.27% reminder 812 2,688 30.21%
sign 1,738 1.23% notion 1,655 5,713 28.97%
notion 1,655 1.17% coincidence 627 2,196 28.55%
point 1,511 1.07% speculation 1,922 6,778 28.36%
warning 1,460 1.03% supposition 46 164 28.05%
fear 1,432 1.01% impression 2,279 8,206 27.77%

As Table 5 shows, the rank list for attraction is dominated by fairly com-
mon, i.e. more or less frequent, nouns. The reason for this is that the overall
frequency of a noun in the corpus does of course have an effect on the like-
lihood of its occurring in any construction. From a purely statistical point of
110 Hans-Jörg Schmid

view, frequent nouns have a better chance than less frequent ones. In con-
trast, the formula used for calculating reliance takes the total frequency of a
noun in the corpus into consideration. As a result, the rank list is headed by
fairly infrequent nouns which, however, are highly specialized, so to speak,
for occurrence in the given pattern. Intuitively, the semantic affinity be-
tween these nouns and the constructions seems to be particularly strong.
A second way of exploiting the notion of reliance is to provide reliance
profiles for individual nouns. These give information on the recurrent colli-
gations entered into by a given noun. Table 6 collects a small number of
examples:
Table 6. Reliance profiles for the nouns idea, finding and temerity

Noun N-to N-BE- N-Th N-BE- N-Wh N-BE- Th-N Th- Freq. in Compiled
to Th Wh BE-N corpus reliance

idea 1271 1141 4812 790 752 13 1674 325 46,654 23.10%
finding 96 32 254 7 586 66.38%
temerity 118 160 73.75%

Legend: N-to = N+to-infinitive; N-BE-to = N+BE +to-infinitive; N-Th = N+that-clause; N-BE-

Th = N+BE+that-clause; N-Wh = N+wh-clause; N-BE-Wh = N+BE+wh-clause; Th-N = demon-
strative determiner+N; Th-BE-N = demonstrative pronoun+BE+N

Table 6 includes absolute scores for frequency of occurrence in the four

constructions focused on in this paper as well as in four others in which
shell nouns are typically found, two containing wh-clauses and two contain-
ing demonstrative determiners or pronouns respectively with anaphoric
function: N + wh-clause (the reason why …), N + BE + wh-clause (the
question is why …), this/that + N (this problem …) and this/that + BE + N
(that’s the problem …). The table shows that idea is a highly versatile noun
that was found to occur in all eight constructions investigated. However, its
score for compiled reliance in the four patterns is below 25%, which means
that not even a fourth of its occurrences in the corpus were found in the
patterns. Finding and temerity, on the other hand, boast fairly high scores
for compiled reliance, but are less versatile or, to put it more positively,
show a much stronger affinity with individual constructions: finding is
primed (cf. Hoey 2005) for occurrence in anaphoric uses and, to a lesser
extent, with that-clauses, while temerity was only found to occur in the
pattern N + to-infinitive, but with a very high reliance score of almost
three-quarters of its 160 total instances in the corpus.
 Does frequency in text instantiate entrenchment? 111

The attraction-reliance method thus provides a way of gauging the re-

ciprocal interaction between nouns and constructions. It captures to some
extent the intuition that some nouns are more important for certain con-
structions than others, and that some constructions are more important for
certain nouns than others. As we will see in Section 7, however, the method
has a number of shortcomings with regard both to its rather crude arithme-
tic and to the interpretation of the output it produces. As new statistical
tools for assessing the attraction of lexemes by constructions have been
proposed since the publication of Schmid (2000), it will be worth looking
into these more advanced statistical techniques before we reflect on the
significance of frequency in Section 7.

5.2 Less simple arithmetic: Collostructional Analysis

In a series of papers, Stefanowitsch and Gries introduced a set of so-called

“collostructional” methods designed to capture in quantitative terms the
5
mutual attraction of lexemes and constructions. Unlike the attraction-
reliance method described in Section 5.1, the collostructional techniques do
not simply rely on counts of observed frequencies. Instead they measure the
degree of likelihood that the patterns of observed frequencies are due to
chance. This can be done by comparing observed frequencies to expected
frequencies, which can be calculated using additional scores derived from
the corpus.
As the following quotation shows, the test case at hand lends itself very
readily to what is known as Collostructional Analysis:
Collostructional analysis always starts with a particular construction and in-
vestigates which lexemes are strongly attracted or repelled by a particular
slot in the construction (i.e. occur more frequently or less frequently than
expected). (Stefanowitsch and Gries 2003: 214)
The actual measure chosen to gauge the degree of attraction is the p-value
6
of a statistical test known as Fisher-Exact. Technically speaking, given a
certain set of observations in a corpus, the p-value indicates the probability
of obtaining this distribution or a more extreme one, assuming the ‘zero-
hypothesis’ that the distribution was the result of chance. Couched in
everyday terms, and as applied to collostructions by Stefanowitsch and
Gries, the smaller the p-value, the higher the probability that the observed
112 Hans-Jörg Schmid

distribution is not due to coincidence and the higher the strength of the
association between lexeme and construction.
Four frequency scores are needed to calculate expected frequencies of
lexemes (L) and constructions (C), as well as p-values (Gries and Ste-
fanowitsch 2003: 218):

1. the frequency of L in C,
2. the frequency of L in all other constructions,
3. the frequency of C with lexemes other than L, and
4. the frequency of all other constructions with lexemes other than L.

The typical output of the test is a list of ‘collexemes’ of a construction to-

gether with ‘their’ p-values indicating the degree of association. More often
than not p-values are so small that their significance resides in the number
7
of decimal places, usually expressed as scores to the power of minus x. To
simplify things, a logarithmic transformation of these scores can be given,
which indicates the number of decimal places. The score illustrated in Note
7 would then simply read ‘20’.
Attractive as this method is, it is not without its pitfalls. While discuss-
ing these shortcomings is beyond the scope of this paper (see Kilgarriff
2005 as well as Schmid and Küchenhoff forthc. for a more detailed cri-
tique), one serious hurdle for unbiased applications must be mentioned
here: the problem of how to determine the score numbered 4 in the list
above (i.e. the frequency of all other constructions with lexemes other than
L). This frequency score serves as a mathematical reference point which is
necessary for calculating the expected frequencies in the 2-by-2 contin-
gency tables serving as input to Fisher-Exact (or other, simpler zero-
hypothesis tests such as the more familiar Chi-square test). However, this
decision is not simply a mathematical but, more importantly, a linguistic
one. The only passage where Stefanowitsch and Gries explicitly address
this problem occurs in connection with the construction ‘N is waiting to
happen’ (2003: 218):
the total number of constructions was arrived at by counting the total num-
ber of verb tags in the BNC, as we are dealing with a clause-level construc-
tion centering around the verb wait.
What this quotation clearly indicates is that in order for the Fisher-Exact
test to make sense linguistically, and not just mathematically, it is necessary
that the construction investigated and the total number of constructions be
paradigmatically related. In a sense, the ‘total number of constructions’
 Does frequency in text instantiate entrenchment? 113

gives the number of constructions which could potentially also have occur-
red instead of the construction under investigation. But this paradigmatic
relation is not unproblematic. For one thing, the constructions under inves-
tigation only occur in the progressive form, so it would have made sense to
choose only verbs in the progressive form as reference constructions. Fur-
thermore, is waiting to happen is a fairly specific type of construction con-
sisting of a verb complemented, or at least followed, by a to-infinitive, and
this again might have called for a more narrowly defined type of reference
construction.
Analogous problems arise in the application of the collostructional
method to the nominal constructions serving as case study here. At first
sight, two extreme choices suggest themselves as solutions: one would be
to use the total number of noun tags in the corpus (ca. 60,000,000); the
other extreme would be to insert only the number of other occurrences of
shell-content constructions (i.e. 422,509 minus the number of tokens of the
intended type). The latter choice would have the advantage of emphasizing
the strong paradigmatic relations in this system, but neglects the fact that
other nouns or nominal constructions could occur instead of shell nouns.
As it turns out, neither of these choices is particularly satisfying. If a
score of 60 million, representing all nouns in the corpus, is entered in the
formula for Fisher-Exact, the calculations will be so demanding that they
go way beyond the capacity of normal computing systems, thus yielding a
p-value of 0 (i.e. infinite likelihood). There is not just a problem with com-
puting power, however, but also one related to the nature of statistical sig-
nificance testing, as an increase in the size of the sample, i.e. the corpus,
investigated also raises the degree of confidence that the differences be-
tween observed and expected frequencies are significant and robust, thus
rendering even arbitrary associations significant (Kilgarriff 2005: 266).
Using the score of 422,509 minus x, on the other hand, does not seem to do
justice to the substantial size of the total reference corpus, which, after all,
provides many more opportunities for constructions comparable to shell-
content constructions to occur.
In view of these difficulties, I have decided to use two different refer-
ence scores in applications of the collostructional method in this paper. One
is the score 422,509 minus x, because this score at least seems to have
some kind of linguistic justification. For a second reference score, the com-
pletely arbitrary number of 10,000,000 was chosen, since it was large en-
ough to reflect the massive size of the corpus but is still manageable to
some extent as regards capacity. While the choice of an arbitrary number
114 Hans-Jörg Schmid

may seem rather odd, from a statistical point of view it is no problem as

long as the same score is used for all lexical items tested in one construc-
tion. As the application of p-values as a measure of attraction strength is
controversial anyway (Stefanowitsch and Gries 2003: 239; cf. Schmid and
Küchenhoff forthc.), and as, therefore, the ranking of items is much more
important than the actual size of the p-value, there is not much to be said
against such a procedure. Table 7 lists the 10 top-ranking nouns each for
attraction and reliance from Table 5 above and gives their p-values for both
reference scores. The nouns are ordered according the p-values in the col-
umn on the far right.
Table 7. Attraction, reliance and p-value scores for selected nouns in the construc-
tion N + that-clause.
p-value,
reference p-value, refer-
Freq. in Freq. in score ence score
pattern corpus Attraction Reliance 10,000,000 281,033
fact 26,106 68,472 18.45% 38,13% 0 0
evidence 5,007 34,391 3.54% 14,56% 0 0
idea 4,812 46,654 3.40% 10,31% 0 0
view 3,532 37,468 2.50% 9,43% 0 0
hope 2,727 16,663 1.93% 16,37% 0 0
news 2,572 49,736 1.82% 5,17% 0 0
feeling 2,511 14,392 1.77% 17,45% 0 0
possibility 2,232 12,075 1.58% 18,48% 0 3,48E-276
doubt 4,010 17,322 2.83% 23,15% 0 2,71E-166
realization 820 1,185 0,58% 69,20% 0 3,01E-139
belief 3,696 9,344 2.61% 39,55% 0 1,52E-41
assumption 1,391 3,151 0,98% 44,14% 0 1,45E-36
impression 2,279 8,206 1.61% 27,77% 0 4,39E-25
assertion 596 1,492 0,42% 39,95% 0 1,25E-07
insistence 796 2,069 0,56% 38,47% 0 1,14E-06
proviso 111 250 0,08% 44,40% 2,95E-134 0,00036
suggestion 2,033 5,854 1,44% 34,73% 0 0,0118
premise 274 765 0,19% 35,82% 5,39E-297 0,16743
misapprehension 44 123 0,03% 35,77% 3,94E-49 0,6329
dictum 84 249 0,06% 33,73% 9,92E-90 0,95
 Does frequency in text instantiate entrenchment? 115

The juxtaposition of the two systems allows for a number of interesting

observations. Firstly, at least with the online statistics lab used for calculat-
ing the Fisher-Exact test, it is impossible to capture differences in associa-
tion strength for a considerable number of the nouns included in the table.
This is true for calculations with either reference score, and is due to the
large size of the corpus used and the resulting high scores for total fre-
quency reached by many nouns (cf. Gries 2005: 278–279, Kilgarriff 2005:
272–273). Secondly, comparatively high p-values close to 1, which indicate
low strengths of attraction, are produced by the test especially for nouns
with a fairly low overall frequency of occurrence, even if their reliance
scores are quite high (cf. the scores for proviso, premise, misapprehension
and dictum). Thirdly, high-frequency nouns with rather low reliance scores
such as view or news leave the test with the same score, i.e. 0, as high-
frequency nouns with much higher reliance scores (e.g. fact) (see Section
6.3 below). It will be useful to keep these observations in mind when we
now turn to a discussion of the cognitive aspects of frequency counts in
corpora.

6. Does frequency really instantiate entrenchment?

6.1 Background
8
In line with the terminological decisions made in Schmid (2007b), the no-
tion of entrenchment is defined as “the degree to which the formation and
activation of a cognitive unit is routinized and automated”. Within Cogni-
tive Linguistics, both this notion of entrenchment and the idea that en-
trenchment correlates with frequency of occurrence can be traced back to
Langacker. According to him, there is a
continuous scale of entrenchment in cognitive organization. Every use of a
structure has a positive impact on its degree of entrenchment, whereas ex-
tended periods of disuse have a negative impact. With repeated use, a novel
structure becomes progressively entrenched, to the point of becoming a
unit; moreover, units are variably entrenched depending on the frequency of
their occurrence. (Langacker 1987: 59)
As this indicates, Langacker conceives of entrenchment as being fostered
by repetitions of cognitive events, i.e. by “cognitive occurrences of any
degree of complexity, be it the firing of a single neuron or a massive hap-
pening of intricate structure and large-scale architecture” (1987: 100). This
116 Hans-Jörg Schmid

seems highly convincing, not least in view of the considerable body of evi-
dence from psycholinguistic experiments suggesting that frequency is one
major determinant of the ease and speed of lexical access and retrieval,
alongside recency of mention in discourse (cf., e.g., Sandra 1994: 30–31,
Schmid 2008, Knobel, Finkbeiner and Caramazza 2008). As speed of ac-
cess in, and retrieval from, the mental lexicon is the closest behavioural
correlate to routinization, this indeed supports the idea that frequency and
entrenchment co-vary.
Nevertheless, it is not easy to transfer Langacker’s idea of ‘massive
happenings of intricate structure’ to larger and complex linguistic units,
since it does not seem to take into consideration that the different compo-
nents of complex linguistics structures may in fact activate each other. For
example, it is not unlikely that shell nouns may trigger their recurring shell
content clauses, or that the clauses may trigger certain shell nouns. The
firing Langacker is talking about may therefore not take place in one go but
rather in a cascade-like fashion, with one element triggering one or more
other elements.
Another problem with Langacker’s view is that he apparently conceives
of frequency in a vacuum. However, as Geeraerts, Grondelaers, and
Bakema (1994) argue, it is not frequency of use as such that determines
entrenchment, but frequency of use with regard to a specific meaning or
function, in comparison with alternative expressions of that meaning or
9
function. Like Brown (1965: 321), Rosch (Rosch et al. 1976: 435) and
Downing (1977: 476) before them, Geeraerts, Grondelaers and Bakema
pursue lexicological rather than grammatical goals and investigate the rela-
tion between the privileged basic level of categorization and the frequency
with which objects are named with terms on this level as opposed to more
general superordinate or more specific subordinate terms. This relative
frequency is indicative of what they call “onomasiological salience”.
As already mentioned in the introduction to this paper, the “From-
corpus-to-cognition principle” somewhat daringly proposed in Schmid
(2000: 39) was inspired by Halliday’s claim that “frequency in text instan-
tiate[s] probability in the [linguistic HJS] system” (Halliday 1993: 3).
Partly responding to legitimate objections that the implications of the from-
corpus-to-cognition principle were far from clear (cf. Esser 2002: 208),
Mukherjee takes up the catchphrase from corpus to cognition and tries to
refine it (cf. Mukherjee 2005: 67, 91, 247): “From a cognitive point of
view, frequency in usage should be best regarded as a quantitative signpost
of the degree of entrenchment” (2005: 225). More precisely, what fre-
 Does frequency in text instantiate entrenchment? 117

quency counts in a corpora reflect more or less directly are degrees of con-
ventionalization of linguistic units or structures. Conventionalization, how-
ever, is a process taking place first and foremost in social, rather than cog-
nitive, systems, and it requires an additional logical step to assume that
degrees of conventionalization more or less directly translate into degrees
of entrenchment. The crucial link of course is frequency of usage and expo-
sure, which on the one hand reflects degrees of conventionalization in the
speech community and on the other hand enhances entrenchment in indi-
10
vidual minds (see Schmid forthc. for more details).
All these attempts to correlate frequency with entrenchment have two
things in common: they presuppose rather than explicitly question know-
ledge about the nature of frequency and they treat frequency as a mono-
lithic concept. More or less the same goes for the considerable body of
literature in grammaticalization theory that tries to relate the frequency of
11
linguistic units to their propensity to grammaticalize. One notable excep-
tion, which will be taken up in the next section, is Hoffmann’s (2004) paper
on the grammaticalization of low-frequency complex prepositions, which
emphasizes the need to be clearer about “what exactly is meant by ‘fre-
quency’” and “what is the relationship between frequency and salience”
(2004: 189).

6.2 Types of frequency

Hoffmann (2004) distinguishes two kinds of frequency, one with two sub-
types. The first type is called conceptual frequency and is reminiscent of
Geeraerts, Grondelaers and Bakema’s notion of onomasiological salience
mentioned in Section 6.1. As Hoffmann notes (2004: 190), this type is dif-
ficult to come to grips with. Its operationalization would require knowledge
of the full range of paradigmatic competitors with regard to one function
and/or meaning. While this is possible in the lexicon it seems hardly viable
to take into consideration all alternative ways of linguistically encoding the
function served by a particular lexico-grammatical construction. Since, at
least in this respect, grammar – and discourse – are much more open-ended
than the lexicon, it does not appear feasible and fruitful to pursue concep-
tual frequency any further in the present study.
The second type is called lexical or textual frequency and is further sub-
divided into absolute and relative frequency. Hoffmann leaves no doubt as
to which of the two he finds more important for corpus-linguistic studies:
118 Hans-Jörg Schmid

[F]requency information for an individual linguistic item only becomes

meaningful as a diagnostic tool if it is compared with the frequency of oc-
currence of related linguistic phenomena. (Hoffmann 2004: 190)
Like Langacker, however, Hoffmann focuses in particular on what Krug
(2003) calls “string frequency” and takes little notice of the possibility of
assessing the frequency of one component of a construction, say the verb in
the ditransitive construction or the shell noun in a shell-content construc-
tion, in relation to another to which it is syntagmatically, rather than para-
digmatically, related (i.e. the complements). Quite clearly this type of ‘rela-
tive’ frequency differs from the one envisaged by Hoffmann. It is therefore
necessary to adapt Hoffmann’s classification to the needs of this study.
In line with Hoffmann’s proposal, the first type can be called absolute
frequency, even though it will still be measured as the relative frequency of
occurrence of a linguistic phenomenon in a given corpus. This is the only
feasible way of operationalizing absolute frequency, since even this meas-
ure needs some kind of reference score and has to be quantified. With re-
gard to shell-content constructions, it is possible to measure the absolute
frequency of six types of linguistic entities in a corpus:
– Absolute frequency of 1) tokens and 2) types of nouns
– Absolute frequency of 3) tokens and 4) types of complements
(‘shell contents’)
– Absolute frequency of 5) tokens and 6) types of constructions (i.e.
nouns in patterns)
The second type of frequency is relative frequency (defined in a way dif-
ferent from Hoffmann’s). Relative frequency can be approached from the
two complementary perspectives introduced in Section 5:
– Attraction: the relative frequency of tokens of noun type vis-à-vis
the frequency of tokens of construction types.
– Reliance: the relative frequency of tokens of noun type in a con-
struction vis-à-vis tokens of the same noun type in other construc-
tions.
Seen from the perspective of the noun, the two types of relative frequencies
are relative to other nouns occurring in the same construction (attraction)
and relative to occurrences of the same noun in other constructions (reli-
ance).
With these distinctions in place we are now in a position to examine the
relations between different types of frequency and degrees of entrenchment
 Does frequency in text instantiate entrenchment? 119

and evaluate the attempts sketched in Section 5 and 6 to quantify this rela-
tion.

6.3 Types of frequency, degrees of entrenchment and arithmetic modelling

In view of the overwhelming evidence in cognitive (neuro-)psychology (see

Section 6.1) it seems safe to accept that repeated patterns of neuronal ac-
tivity foster the entrenchment, routinization and activation of the corres-
ponding cognitive events. If this seems plausible enough, then it will also
make sense to acknowledge that the absolute frequency of occurrence of
types of linguistic entities will show a relationship to the degree of en-
trenchment of their cognitive and neurological correlates (whatever these
may be). This follows from the assumption that frequency of occurrence in
discourse relates to frequency of processing in the minds of the members of
12
the speech community.
This is not the whole story, however. In fact, psycholinguistic evidence
also exists which suggests that this relation may apply to linguistic forms
irrespective of their function and meaning. Thus, in a classic study,
Swinney (1979) demonstrated that during lexical access, i.e. roughly the
first third of a second after being confronted with a word-form, test subjects
activate both contextually appropriate and contextually inappropriate mean-
ings of homonyms such as bug (‘insect’ vs. ‘overhearing device’). More
recent production experiments have even suggested that low-frequency
forms (e.g. nun) profit with regard to their speed of activation from the
existence of high-frequency homophones (none), and this in spite of the
fact that the two forms represent two different lexemes whose meanings are
totally unrelated (Jescheniak and Levelt 1994, Jescheniak, Meyer and
13
Levelt 2003). The frequency determining the ease of lexical access may
possibly not be the word-specific frequency (e.g. of nun as opposed to
none) but the cumulative frequency of all the homophonic forms (i.e. fre-
quency of nun plus frequency of none).
This finding from lexical access studies can be transferred to the prob-
lem at hand. If it is true that even homophonic (but not homographic) forms
influence each other with regard to entrenchment, then it would also seem
very likely that different usage-patterns of one and the same noun lead to
cumulative entrenchment. This in turn would suggest that the overall token
frequency of nouns in the corpus (in all environments) will have an effect
on their entrenchment, both in a certain shell-content construction and in all
120 Hans-Jörg Schmid

other environments in which they occur. For example, high-frequency

nouns like time, point or way are most likely more entrenched than less
frequent ones (like disinclination or unwillingness), irrespective of their
actual linguistic environment. In stark contrast to received opinion in state-
of-the-art (cognitive) corpus linguistics, epitomized for instance in the pas-
sage from Hoffmann (2004) quoted in the previous section, this means that
there is after all an absolute, cotext-free type of entrenchment, which corre-
lates with absolute frequency of occurrence.
The question now is whether cotext-free entrenchment is integrated in
quantitative accounts of attraction strengths. In the attraction-reliance
framework introduced in Section 5.1, absolute frequency is included as a
factor, albeit tacitly rather explicitly, since the formula used for calculating
attraction scores does not include the overall frequency of a given noun in
the corpus. As a result, the rank lists for attraction tend to be headed by
nouns with fairly high absolute frequencies. Mathematically ill-informed as
this clearly is, it may in fact have a certain degree of cognitive plausibility,
as it allows absolute frequency scores to influence and even supersede rela-
tive frequencies in the patterns. In the collostructional framework (cf. Sec-
tion 5.2), absolute frequency constitutes an integral part of the calculation
of p-values, since it is entered in the contingency tables and thus automati-
cally and deliberately taken into account. The scores given in Section 5.2
even suggest that the test exaggerates the effect of absolute frequency
mathematically, as higher absolute frequencies increase the confidence in
the assessment of the dataset and thus automatically result in lower p-
values, i.e. higher attraction scores.
If absolute frequency translates into the cognitive system as cotext-free
entrenchment, it seems reasonable to think of its relative counterpart as
reflecting cotextual entrenchment. Very much in line with Hoey’s (2005)
idea of lexical priming, cotextual entrenchment can be seen as the tendency
of one linguistic element or unit to trigger the (co-)activation of one or
more other linguistic units or structures in language users’ minds, if the
former significantly co-occurs with the latter in actual discourse. Elements
co-occurring frequently are intuitively held to be more cotextually en-
trenched vis-à-vis each other than elements rarely found in each other’s
company.
The trouble with cotextual entrenchment is that, as we have just seen, its
strength will inevitably be influenced by the effects of cotext-free en-
trenchment. Still worse, the strength of this effect is difficult to gauge.
Theoretically, the full range of combinations of cotext-free and cotextual
 Does frequency in text instantiate entrenchment? 121

entrenchment are possible. To facilitate further discussion of the interaction

of cotext-free and cotextual entrenchment, combinations of extreme values
are cross-tabulated in Table 8. They are illustrated with hand-picked every-
day examples as well as typical cases of shell-content constructions (which
are accompanied by relative frequency scores for occurrence in the respec-
tive pattern and absolute frequency scores in the corpus).
Table 8. Theoretical combinations of extremes of cotext-free and cotextual en-
trenchment

Cotext-free entrenchment
high low
get up; with kith and kin;
high fact + that-clause disinclination + to-infinitive
Cotextual (26,106 out of 68,472) (45 out of 62)
entrenchment get low; shopgrift a nouse;
low way + BE + to-infinitive aphorism + BE + that
(316 out of 201, 366) (1 out of 81)

The bottom right-hand cell is undoubtedly the one presenting the fewest
problems. In everyday terms, this cell describes the occurrence of rare
words in uncommon uses. So far the verb shopgrift (“the activity of pur-
chasing something from a shop, using it, and then returning it within a spe-
cific period in order to get a full refund”, Maxwell 2006, s.v. shopgrifting)
and the noun nouse (“a pointing mechanism for a personal computer which
is activated by movements of the nose”, Maxwell 2006, s.v. nouse) are
1
hardly established neologisms with a low frequency of occurrence. In addi-
tion, their combination is odd, to say the least. In a similar vein, aphorism
is a fairly rare noun in the COBUILD corpus and was found to occur only
once in the pattern N + BE + that-clause. While the noun itself may well be
entrenched in some people’s minds, for example literary scholars or teach-
15
ers of rhetoric, it can hardly be considered a salient lexeme, neither in this
construction nor elsewhere.
Similarly straightforward, but complementary cases are captured in the
top left-hand cell. There can be no doubt that fact is deeply entrenched in
most adult speakers’ minds. The high proportion of uses in the pattern N +
that-clause (26,106 out of a total of 68,472) also predicts a high level of
cotextual entrenchment. In the attraction-reliance framework, this is re-
flected in a combination of high scores for both attraction and reliance. In
the collostructional framework, cases like these are the best candidates for
producing p-values of 0, which loosely speaking indicates an infinitely high
122 Hans-Jörg Schmid

probability that the observed frequencies are not due to chance. In cases of
this type, it is a moot point whether the strong scores for attraction are a
result of superseding absolute frequency or due to relative frequency. Put
rather bluntly, the noun fact is entrenched, the N + that-clause construction
is entrenched, and the lexically filled construction fact + that-clause is en-
trenched as well, just as the verb get, the particle up and the phrasal verb
get up are entrenched.
The top right-hand cell is a bit more problematic. Kith is a very rare lin-
guistic form; therefore it is very likely that it rates low with regard to
cotext-free entrenchment. With regard to cotextual entrenchment in the
fixed expression with kith and kin, however, it clearly rates high, since it
has no other habitat to thrive in. Here cotextual entrenchment clearly comes
to the fore, as it is not influenced by cotext-free entrenchment. However,
extreme cases of this type are more or less restricted to the domain of
phraseology. The closest approximation in the area of shell-content con-
structions is found for low-frequency nouns relying heavily on the pattern
N + to-infinitive for the occurrence in discourse. The noun disinclination,
which boasts a reliance score of 73.75% (45 out of 62), is a case in point.
Intuitively, examples of this type show the highest degree of semantic af-
finity with the matching pattern, and thus also of cotextual entrenchment. In
the attraction-reliance method, however, this strong affinity only shows up
in the reliance scores; in the rank list for attraction, which ranks the 200
nouns found in the pattern N + to-infinitive for their frequency in that pat-
tern, disinclination occupies rank 193. In the Fisher-Exact test, the p-values
for cases of this type are astonishingly high, indicating a comparatively low
attraction score. The combination disinclination + to-infinitive yields p-
values of 2.42e-60 (for a reference score of 10 million) and 0.00321 (for a
reference score of 194,244, i.e. all valid tokens of shell-content construc-
tion minus the 228,165 tokens of the N + to-infinitive construction). The
presumably high degree of cotextual entrenchment is not reflected particu-
larly in the second score because – as discussed in Section 6 – low absolute
frequencies reduce the confidence of the Fisher-Exact test.
Finally, the bottom left-hand cell is where we can observe how absolute
entrenchment can get the better of relative entrenchment. No more than 316
out the mass of 201,366 tokens of the noun way in the corpus were found in
the construction N + BE + to-infinitive, most of them in the more specific
patterns the only way is/was to … and the best way is/was to …. Now, while
these patterns do sound familiar and are thus most likely cotextually en-
trenched in most speakers’ minds, they are neither typical instantiations of
 Does frequency in text instantiate entrenchment? 123

the noun nor of the construction. The construction has a much stronger
association with mental and deontic nouns like aim, intention, ambition,
task and job (cf. Schmid 2007a for a semantic analysis of this construction).
Here my feeling would be that the enormous cotext-free entrenchment of
the semantically highly unspecific noun way, which lends itself to uses in a
huge range of different patterns, clearly overrides its cotextual entrench-
ment (which cannot be ignored, however). Do the two quantitative methods
capture this effect? The attraction-reliance method lists way as ranking 20th
(out of 162 types) in terms of attraction, and 135th in terms of reliance. In a
sense, this combination of ranks reflects our intuition concerning the effects
of cotext-free and cotextual entrenchment, but it does not help a lot in actu-
ally quantifying them with any degree of precision. In the Fisher-Exact test,
the p-values for way + BE + to-infinitive are in a way complementary to
those obtained for disinclination + to-infinitive: while the latter yielded a
decently low p-value for a reference score of 10 million and a fairly large
one for the closed-system reference score, way yields the score 0 for the
smaller reference score of 400,624, and 4.86e-9 for the larger one.
Where do we stand now? This comparison of hypothetical patterns of
entrenchment and two different attempts to capture them quantitatively has
shown that we seem to be quite far from having a good grip on the relation
between frequency and entrenchment. This is mainly due to the unclear
interaction between absolute and relative frequency, or cotext-free and
cotextual entrenchment, respectively. While some patterns of this interac-
tion as manifested in observed frequencies may in fact be reflected quite
well in the scores for attraction and reliance, the attraction-reliance method
may be criticized for being unable to produce one single score for en-
trenchment that takes both cotext-free and cotextual entrenchment into
account. Even if we accept that the reciprocal attraction of constructions
and nouns is a two-dimensional phenomenon that deserves two measures, it
still remains a problem that the method works with raw, observed frequen-
cies and does not include any tests of significance. The more sophisticated
collostructional method, on the other hand, does exploit statistical tests of
significance, relates observed to expected frequencies and also takes abso-
lute frequency into consideration (though possibly with exaggerated ef-
fects). However, its application is seriously impeded by the uncertainty
concerning the appropriate choice of reference scores, which have a strong
effect on the p-values indicating the strength of attraction. Furthermore, the
exclusive reliance in this method on significance testing risks masking im-
portant distributional differences which are very likely also reflections of
124 Hans-Jörg Schmid

different entrenchment patterns. This can be illustrated with the help of the
fictive examples juxtaposed in Table 9 below.

Table 9. Juxtaposition of different fictive frequency distributions and their reliance

scores and p-values
High relative frequency – low absolute Low relative frequency – high absolute
frequency; reliance = 40.00% frequency; reliance = 2.26 %
N in construction Construction with N in construction Construction with
other nouns other nouns
40 22,000 113 21,886

100 10,000,000 5000 99,951,41

N in corpus in Other constructions N in corpus in Other constructions

other construc- with other nouns other constructions with other nouns
tions
p = 6.74E-72 p = 8.78e-72

The left-hand side of the table exemplifies a case of high reliance caused by
the combination of high relative with low absolute frequency (similar to the
type disinclination + to-infinitive, but less extreme). In contrast, the fre-
quency pattern on the right-hand side shows a moderately frequent noun
with a fairly small number of occurrences in the given pattern (reliance
score 2.26%). The reference scores in the right-hand cells of the two col-
umns only differ because the grand total (representing the sum of the scores
in all four cells) must remain stable in order for the comparison to be cor-
rect. The crucial point here is that despite the striking differences in relative
vs. absolute frequency, both patterns produce an almost identical p-value in
the Fisher-Exact test. This means the two fictive nouns would turn out to
have identical attraction strengths to the given construction representing
identical degrees of cotextual entrenchment, which seems somewhat mis-
leading.
What I have not considered so far in this section are all kinds of combi-
nations of more or less medium values for absolute and relative frequen-
cies, presumably reflecting medium degrees of cotext-free and cotextual
entrenchment. While some of the cells included in Table 9 seem to be at
least theoretically straightforward, if we are honest we must admit that we
know very little about how to deal with these ‘mediocre’ cases. It seems
very plausible that combinations with high scores for cotext-free and cotex-
 Does frequency in text instantiate entrenchment? 125

tual entrenchment such as the fact that …, the aim is to …, the problem is
that … and the attempt to … serve metaphorically speaking as conceptual
anchors of the respective construction types. They more or less have the
status of fixed phrases and are most likely retrieved as one chunk, as sug-
gested by Sinclair’s (1991) idiom principle. At the other extreme, combina-
tions of the type the aphorism that … (or even more shopgrift a nouse) may
in fact catch our attention simply because they are so unfamiliar to us – an
effect exploited for rhetorical and stylistic means, e.g. in poetry, journalism
and advertising. Whether the huge bulk of combinations between these
extremes in fact show the kind of linear proportional correlation between
frequency and entrenchment which is usually taken to exist, is an open
question.

7. Conclusion

Unfortunately, but also perhaps not surprisingly in view of the preceding

discussion, this paper has to end on a somewhat less-than-enthusiastic note.
It seems to me that many researchers, inclusing myself, have had a great
deal too much confidence in the potential of quantitative methods for the
study of aspects of the linguistic and cognitive system. All quantitative
methods that I am aware of ultimately boil down to counting the frequen-
cies of tokens and types of linguistic phenomena. What I have tried to show
here, however, is that so far we have understood neither the nature of fre-
quency itself nor its relation to entrenchment, let alone come up with a
convincing way of capturing either one of them or the relation between
them in quantitative terms. This remains true in spite of the indisputable
advantages of quantitative methods such as their predictive power, the
possibility to falsify models by means of repeat analysis and their enormous
capacity when it comes to coming to grips with highly multivariate
datasets. Essentially, this failure is caused by the following complications.
Firstly, frequency of occurrence is a much less objective measure than
most proponents of quantitative (cognitive) linguistics seem to realize. The
assessment of frequency scores depends not only on what researchers re-
trieve and count as valid tokens, but also on how they calculate frequency.
Even if they show awareness of the need to distinguish absolute from rela-
tive frequency (as of course most practitioners do), then it is still unclear
how the two interact with each other, since absolute frequency may not be
as irrelevant as most corpus linguists think. Secondly, advanced statistical
126 Hans-Jörg Schmid

techniques, which take absolute frequencies into consideration in order to

gauge the significance of observed relative frequencies, have the problem
of determining the reference scores required for the tests and run the risk of
obscuring different combinations of absolute and relative frequency of oc-
currence. Thirdly, even if we accept the plausibility of the general claim
that frequency of processing, and thus of occurrence in discourse, correlates
with strength of entrenchment, we are still underinformed about the relation
between cotext-free and cotextual entrenchment. This is particularly true of
the large bulk of cases showing a medium range of association of lexeme
and construction. Recent attempts at tallying results from corpus studies
with results from experimental methods, for example by Gries, Hampe and
Schönefeld (2005, 2010) and Wiechmann (2008; cf. also Gilquin and Gries
2009), point to one direction where additional information may be avail-
able. While it must be stressed that psycholinguistic experiments represent
just another way of trying to tap into the black box, whose ‘real’ workings
will remain hidden to us for some time, converging evidence produced by
different methods is undoubtedly superior to results from either corpus or
experimental studies.

Notes

1. I would like to thank Joan Bybee, Susanne Handl, Laura Janda, Adam Kilgar-
riff, Manfred Krug and John Newman for invaluable comments on earlier ver-
sions of this paper. I am also indebted to the participants of the theme session
at the Krakow ICLA conference (July 2007), of the workshop on “Chunks in
Corpus Linguistics and Cognitive Linguistics” Erlangen/Germany (October
2007) and the attendees of a guest lecture at Freiburg University (May 2008)
for their input into this study.
2. It should not go unnoted that in lexicography and descriptive grammar the
relevance of the frequency of patterns was recognized very early by John Sin-
clair and taken into consideration in the design of entries in the first edition of
the COBUILD dictionary (Sinclair et al. 1987) and the first COBUILD grammar
(Sinclair 1990). For an account of further developments in lexicography in the
1990s, see Kilgarriff (1997).
3. In fact, as the recent study by Mukherjee (2005) on ditransitive verbs shows,
the idea that frequency of occurrence relates to (proto-)typicality is – more or
less explicitly – still going strong, despite the debate in the 1980s triggered by
 Does frequency in text instantiate entrenchment? 127

Rosch’s (1975) quantitative approach to prototypicality; cf. Schmid 1993: 27–

28). For a critique of this approach, see Gilquin (2006).
4. See Schmid (2000: 301-376) for more details on the shared semantic, textual
and cognitive functions of shell-content constructions.
5. Cf. Stefanowitsch and Gries (2003), Gries and Stefanowitsch (2004, 2010),
Gries (2006a, 2006b) and Stefanowitsch (2005).
6. The Fisher-Exact test is part of most available statistics programmes such as R
or SPSS, but it can also be found online; see Wulff (2005) for a useful survey
of sites.
7. The scores are usually expressed as, e.g., 2.345e-20, which reads “2.345 to the
power of minus 20”, i.e. 0.00000000000000000002345.
8. In contrast to entrenchment, the notion of salience is not taken to refer to
degrees of routinization, but either to temporary activation states of mental
concepts (referred to as cognitive salience) or to inherent and consequently
more or less permanent properties of entities in the real world (i.e. ontological
salience; cf. Schmid, in print). The relation between the two notions is quite
complex: on the one hand, ontologically salient entities attract our attention
more easily and thus more frequently than nonsalient ones. As a result, cogni-
tive events related to the processing of ontologically salient entities will occur
more frequently and lead to an earlier entrenchment of corresponding cogni-
tive units, i.e. concepts. On the other hand, deeply entrenched cognitive units
are more likely to become cognitively salient than less well entrenched ones,
because a smaller amount of spreading activation will suffice to activate them.
9. As will be shown in Section 7.3, there is psycholinguistic evidence suggesting
that absolute frequency of occurrence may be an important factor after all.
10. In contrast to Langacker (2008: 21, fn.13) I consider it important to keep the
notions of entrenchment and conventionalization apart. As pointed out by
Langacker, entrenchment is a matter of individual minds whereas convention-
ality and conventionalization are notions pertaining to speech communities.
While these two systems, the cognitive and the social, are intricately inter-
twined, they are governed by different kinds of structures and processes: asso-
ciation, chunking, automatization, generalization and categorization in the
cognitive system, as opposed to innovation, accommodation, diffusion and
normation in social systems (cf. Schmid forthc.).
11. An early milestone in this tradition is Bybee (1985). Recent publications look-
ing closer into the relation between frequency, grammaticalization tendency
and entrenchment or salience include Wray (1999), Croft (2000), Bybee
(2001), (2003), (2006), Krug (2003), Hoffmann (2004) and Mair (2004), as
well as the collection of articles edited by Bybee and Hopper (2001).
12. This of course does not imply that highly frequent items or patterns automati-
cally correspond to what are known as prototypes in Cognitive Semantics (cf.
Gilquin 2006), because there are other factors determining prototypicality, e.g.
128 Hans-Jörg Schmid

perceptual salience or conceptual complexity (and because there is no agree-

ment on how to define the notion in the first place).
13. It should be added that the cumulative frequency effect of homophones has
been questioned by other researchers, most notably by Carramazza et al.
(2001) and (2004).
14. Google searches performed on 10 August 2007 produced no more than 79 hits
on English-language pages for the form shopgrift, and 174,000 hits for nouse
(The form mouse yielded 117,000,000 hits on that day).
15. This is a reminder of Hoey’s (2005) important insight that degrees of lexical
priming (and thus cotextual entrenchment) are register- and even speaker-
dependent (cf. Schmid 2007c).

References

Brown, Roger
1965 Social Psychology. New York: Free Press.
Brugman, Claudia M.
1981 The Story of Over. Trier: LAUT.
Bybee, Joan
1985 Morphology. A study of the relationship between meaning and form.
Amsterdam / Philadelphia: John Benjamins.
2001 Frequency effects on French liaison. In Frequency and the Emer-
gence of Linguistic Structure, Joan Bybee and Paul Hopper (eds.),
337-359. Amsterdam / Philadelphia: John Benjamins.
2003 Mechanisms of change in grammaticization: The role of frequency.
In Handbook of Historical Linguistics, Brian D. Joseph and Richard
D. Janda (eds.), 602-623. Oxford: Blackwell.
2006 From usage to grammar: The mind’s response to repetition. Lan-
guage 82: 711-733.
Caramazza, Alfonso, Albert Costa, Michele Miozzo, and Yanchao Bi
2001 The specific-word frequency effect: Implications for the representa-
tion of homophones. Journal of Experimental Psychology: Learning,
Memory, and Cognition 27: 1430-1450.
Caramazza, Alfonso, Yanchao Bi, Albert Costa and Michele Miozzo
2004 What determines the speed of lexical access: Homophone or spe-
cific-word frequency? A reply to Jescheniak et al. (2003). Journal of
Experimental Psychology: Learning, Memory, and Cognition 30:
278-282.
 Does frequency in text instantiate entrenchment? 129

Church, Kenneth W. and Patrick Hanks

1990 Word association norms, mutual information & lexicography. Com-
putational Linguistics 16: 22-29.
Clear, Jeremy
1993 From Firth principles. Computational tools for the study of colloca-
tion. In Text and Technology. In honour of John Sinclair, Mona
Baker, Gill Francis and Elena Tognini-Bonelli (eds.), 271-292.
Amsterdam / Philadelphia: John Benjamins.
Croft, William
2000 Explaining Language Change. An evolutionary approach. Harlow:
Longman.
Dirven, René
1991 Schema and subschemata in the lexical structure of the verb agree.
Cahiers de l'Institut de Linguistique de Louvain 17: 25-42.
Downing, Pamela
1977 On ‘basic levels’ and the categorization of objects in English dis-
course. In Proceedings of the Third Annual Meeting of the Berkeley
Linguistics Society, Kenneth W. Whistler, et al., (eds.), 475–487.
Berkeley: Berkeley Linguistics Society.
Esser, Jürgen
2002 Review of H.-J. Schmid (2000), English Abstract Nouns as Concep-
tual Shells. From corpus to cognition. In Anglistik 13: 204-208.
Geeraerts, Dirk
1983 Prototype theory and diachronic semantics. A case study. Indoger-
manische Forschungen 88: 1-32.
Geeraerts, Dirk, Stef Grondelaers, and Peter Bakema
1994 The Structure of Lexical Variation. Meaning, naming, and context.
Berlin / New York: Mouton de Gruyter.
Gilquin, Gaëtanelle
2006 The place of prototypicality in corpus linguistics. Causation in the
hot seat. In Corpora in Cognitive Linguistics. Corpus-based ap-
proaches to syntax and lexis, Stefan Th. Gries and Anatol Ste-
fanowitsch (eds.), 159–191. Berlin / New York: Mouton de Gruyter.
Gilquin, Gaëtanelle, and Stefan Th. Gries
2009 Corpora and experimental methods: A state-of-the-art review. Cor-
pus Linguistics and Linguistics Theory 5: 1–26.
Goldberg, Adele
1995 Constructions: A Construction Grammar approach to argument
structure constructions. Chicago: University of Chicago Press.
130 Hans-Jörg Schmid

Gries, Stefan Th.

2005 Null-hypothesis significance testing of word frequencies: A follow-
up on Kilgarriff. Corpus Linguistics and Linguistic Theory 1: 277-
94.
Gries, Stefan Th. and Anatol Stefanowitsch
2004 Extending collostructional analysis: A corpus-based perspective on
‘alternations’. International Journal of Corpus Linguistics 9: 97-
129.
2010 Cluster analysis and the identification of collexeme classes. In Em-
pirical and Experimental Methods in Cognitive/Functional Re-
search, Sally Rice and John Newman (eds.), 73-90. Stanford, CA:
CSLI.
Gries, Stefan Th., Beate Hampe and Doris Schönefeld
2005 Converging evidence: Bringing together experimental and corpus
data on the associations of verbs and constructions. Cognitive Lin-
guistics 16: 635-676.
2010 Converging evidence II: More on the association of verbs and con-
structions. In Empirical and Experimental Methods in Cogni-
tive/Functional Research, Sally Rice and John Newman (eds.), 59-
72. Stanford: CSLI.
Halliday, M.A.K.
1993 Quantitative studies and probabilities in grammar. In Data, Descrip-
tion, Discourse. Papers on the English language in honour of John
McH. Sinclair, Michael Hoey (ed.), 1-25. London: Harper Collins.
Herskovits, Anna
1986 Language and Spatial Cognition: An interdisciplinary study of the
prepositions in English. Cambridge: Cambridge University Press.
Hoey, Michael
2005 Lexical Priming. A new theory of words and language. London and
New York: Routledge.
Hoffmann, Sebastian
2004 Are low-frequency complex prepositions grammaticalized? On the
limits of corpus data – and the importance of intuition. In Corpus
Approaches to Grammaticalization in English, Hans Lindquist and
Christian Mair (eds.), 171-210. Amsterdam / Philadelphia: John
Benjamins.
Jescheniak, Jörg D. and Willem J. M. Levelt
1994 Word frequency effects in speech production: Retrieval of syntactic
information and of phonological form. Journal of Experimental Psy-
chology: Learning, Memory, and Cognition 20: 824-843.
 Does frequency in text instantiate entrenchment? 131

Jescheniak, Jörg D., Antje Meyer and Willem J. M. Levelt

2003 Specific-word frequency is not all that counts in speech production:
Comments on Caramazza, Costa, et al. (2001) and new experimental
data. Journal of Experimental Psychology: Learning, Memory, and
Cognition 29: 432-438.
Kilgarriff, Adam
1997 Putting frequencies in the dictionary. International Journal of Lexi-
cography 10: 135-155.
2005 Language is never, ever, ever, random. Corpus Linguistics and Lin-
guistic Theory 1(2): 263-276.
Knobel, Mark, Matthew Finkbeiner and Alfonso Caramazza
2008 The many places of frequency: Evidence for a novel locus of the
lexical frequency effect in word production. Cognitive Neuropsy-
chology 25: 256-86.
Krug, Manfred
2003 Frequency as a determinant in grammatical variation and change. In
Determinants of Grammatical Variation in English, Günter Rohden-
berg and Britta Mondorf (eds.), 7–67. Berlin / New York: Mouton
de Gruyter.
Lakoff, George
1982 Categories and Cognitive Models. Trier: LAUT.
1987 Women, Fire, and Dangerous Things. What categories reveal about
the mind. Chicago / London: University of Chicago Press.
Langacker, Ronald W.
1987 Foundations of Cognitive Grammar, Vol. 1. Theoretical prerequi-
sites. Stanford: Stanford University Press.
Lindner, Susan J.
1982 A lexico-semantic analysis of English verb particle constructions
with out and up, PhD Dissertation, San Diego, University of Cali-
fornia.
Mair, Christian
2004 Corpus linguistics and grammaticalization theory: Statistics, fre-
quencies, and beyond. In Corpus Approaches to Grammaticalization
in English, Hans Lindquist and Christian Mair (eds.), 121-150.
Amsterdam / Philadelphia: John Benjamins.
Matthiessen, Christian M.I.M.
2006 Frequency profiles of some basic grammatical systems: An interim
report. In System and Corpus: Exploring connections, Geoff
Thompson and Susan Hunston (eds.), 103-142. London: Equinox.
Maxwell, Kerry
2006 From al desko to zorbing. New words for the 21st century. London:
Pan MacMillan.
132 Hans-Jörg Schmid

Mukherjee, Joybrato
2005 English Ditransitive Verbs. Aspects of theory, description and a
usage-based model. Amsterdam / New York: Rodopi.
Rosch, Eleanor
1975 Cognitive representations of semantic categories. Journal of Ex-
perimental Psychology 104: 193–233.
Rosch, Eleonor, Caroline B. Mervis, Wayne D. Gray, David M. Johnson, and Penny
Boyes-Braem
1976 Basic objects in natural categories. Cognitive Psychology 8: 382–439.
Rudzka-Ostyn, Brygida
1989 Prototypes, schemas, and cross-category correspondences: The case
of ask. Linguistics 27: 613-661.
Sandra, Dominiek
1994 Morphology in the reader’s mental lexicon. Frankfurt/Main: Peter
Lang.
Schmid, Hans-Jörg
1993 Cottage and Co., Idea, Start vs. Begin. Die Kategorisierung als
Grundprinzip einer differenzierten Bedeutungsbeschreibung. Tübin-
gen: Niemeyer.
2000 English Abstract Nouns as Conceptual Shells. From corpus to cog-
nition. Berlin / New York: Mouton de Gruyter.
2007a Non-compositionality and emergent meaning of lexico-grammatical
chunks: A corpus study of noun phrases with sentential comple-
ments as constructions. Zeitschrift für Anglistik und Amerikanistik
3(3): 313-340
2007b Entrenchment, salience and basic levels. In The Oxford Handbook of
Cognitive Linguistics, Dirk Geeraerts and Hubert Cuyckens (eds.),
117-138. Oxford: Oxford University Press.
2007c Review of Hoey (2005). Lexical priming. A new theory of words and
language. London: Routledge. In Anglia 125(2): 339–342.
2008 New words in the mind. Concept-formation and entrenchment of
neologisms. Anglia 126(1): 1–36.
forthc. Puzzling over entrenchment and conventionalization. In Construc-
tions – Collocations – Patterns, Thomas Herbst, Hans-Jörg Schmid
and Susen Faulhaber (eds.)
Schmid, Hans-Jörg and Helmut Küchenhoff
forthc. Looking behind the scenes of collostructional analysis.
Schulze, Rainer
1988 A short story of down. In Understanding the Lexicon. Meaning,
sense and world knowledge in lexical semantics, Werner Hüllen and
Rainer Schulze (eds.), 395–414. Tübingen: Niemeyer.
 Does frequency in text instantiate entrenchment? 133

Sinclair, John M.
1990 Collins COBUILD English Grammar. London: Collins.
1991 Corpus, Concordance, Collocation. Oxford: Oxford University
Press.
Sinclair, John M., et al.
1987 Collins COBUILD English Language Dictionary. London: Collins.
Stefanowitsch, Anatol
2005 New York, Dayton (Ohio), and the raw frequency fallacy. Corpus
Linguistics and Linguistic Theory 1(2): 295-301.
Stefanowitsch, Anatol, and Stefan Th. Gries
2003 Collostructions: Investigating the interaction of words and construc-
tions. International Journal of Corpus Linguistics 8(2): 209-243.
Stubbs, Michael
1995 Collocations and semantic profiles. On the cause of the trouble with
quantitative studies. Functions of Language 2: 23-55.
Swinney, David A.
1979 Lexical access during sentence comprehension: (Re)consideration of
context effects. Journal of Verbal Learning and Verbal Behavior 18:
645-659.
Tognini Bonelli, Elena
2001 Corpus Linguistics at Work. Amsterdam / Philadelphia: John Ben-
jamins.
Tummers, Jose, Kris Heylen, and Dirk Geeraerts
2005 Usage-based approaches in Cognitive Linguistics: A technical state
of the art. Corpus Linguistics and Linguistic Theory 1(2): 225-261.
Wiechmann, Daniel
2008 On the computation of collostruction strength: Testing measures of
association as expressions of lexical bias. Corpus Linguistics and
Linguistic Theory 4(2): 253–290.
Wulff, Stefanie
2005 Online statistics labs. Corpus Linguistics and Linguistic Theory 1(2):
303-308.
Section II

Advancing the science: Theoretical questions

The aspectual coercion of the English Durative
Adverbial

Stefan Fuhs

Abstract
This paper investigates the aspectual semantics of the English Durative Adverbial
Construction, such as I talked with him for ages or I tend to worry about it for six
weeks. This construction is a well-known test frame for the telicity of the clause or
phrase with which it is combined, and it is in this regard that it has received thor-
ough discussion in the literature. The study offers a Collostructional Analysis of
this construction, which reports the verbs significantly attracted to clauses contain-
ing the durative adverbial. The set of significantly attracted collexemes and their
aspectual construal in the clausal cores show that the durative adverbial is not
solely restricted to the co-occurrence with collexemes conventionally described as
atelic. Instead, it is argued that the English durative adverbial is just as much part
of the aspectual makeup as any other aspect construction and as such it can lead to
a reconstrual of collexemes so that they are interpretable in the atelic scope of this
construction.
Keywords: aspectuality, Collostructional Analysis, Durative Adverbial Construc-
tion

1. Introduction

This paper investigates the aspectual semantics of the English Durative

Adverbial Construction, such as I talked with him for ages or I tend to
worry about it for six weeks. This construction is well known in the litera-
ture on aspect, mostly for its tendency to ‘select’ for atelic predicates in the
sense that the co-occurrence of the former with the latter is utilized as a
litmus test for telicity. Previous studies of the Durative Adverbial Construc-
tion have often focused on a qualitative, but not a quantitative description
of this construction, thereby leaving a gap in a functionalist and usage-
based exploration of the constructional semantics which takes frequency
bands and association strengths seriously as a part of grammatical descrip-
tion. The present study adopts a Construction Grammar view of grammar
and semantics in which the English durative adverbial is considered to be
138 Stefan Fuhs

an aspect construction, i.e. a construction that influences the aspectual con-

strual of the predicate with which it co-occurs. In Construction Grammar,
constructions have meaning in the same way that lexemes have meaning,
rather than being assigned to different areas of description as manifested by
a theoretical division of lexicon and grammar.
The present study offers an empirical investigation of the constructional
semantics of the English Durative Adverbial Construction and thus contri-
butes new insights to the study of aspectuality in general and the aspectual
semantics of the construction in particular. It addresses a question raised in
the literature on aspect, namely what role the adverbial plays in aspectual
composition and whether it could be treated as detached from the param-
eters that contribute directly to the aspectual type of the clause (cf. Sasse
2001: 257). A study that investigates association strength, hence perform-
ing a quantitative investigation, can illustrate properties of Cognitive Se-
mantics that non-quantitative endeavors could not render visible. Especially
the investigation of the construal of a verb-argument cluster to an adverbial
construction is not so much a matter a possible occurrence but rather of
probable occurrence. In brief, it is shown that the durative adverbial is an
aspect construction that is directly associated with a specific construal of
specific verb stems and it is argued that it should therefore be considered to
be part of the aspectual composition proper. Furthermore, it is argued that
the surprisingly high co-occurrence of the atelic durative adverbial with
telic predicates is an indication that so-called coercion effects are much
more common than it is at least suggested in many treatments of the subject
matter.

2. Aspectuality and previous work on the English Durative

Construction

Any study on aspectuality must address some fundamental issues of termi-

nology and theoretical commitment. Admittedly, this applies to most areas
of inquiry, but especially the study of aspect and aspectuality has seen a
remarkable range of variety across almost all distinguishable parameters. I
will restrain this discussion here to a fairly concise overview in order to
sketch out the framework adopted for this study. For an exhaustive and
recent overview of the state-of-affairs in the study of aspect, the reader is
pointed to a review article by Sasse (2002).
 Aspectual coercion of the English Durative Adverbial 139

Aspect is best defined as the internal temporal constituency of a situa-

tion, following Comrie (1976: 3). In research originating mainy from work
on Slavic languages, but no longer restricted to this domain, aspect is
understood as a morphological category which distinguishes perfective
from imperfective situations. It is also referred to as aspect proper or view-
point aspect (cf. Smith 1997: 2-5). It can be contrasted with lexical aspect,
occasionally still referred to as Aktionsart, which distinguishes the classical
situation types: states, activities, accomplishment and achievements. This
original list of types has been somewhat extended and refined over the
years and it often includes semelfactives (Smith 1997: 3, Sasse 2002: 253).
This bipartite distinction between lexical and grammatical aspect has been
challenged; for instance, Sasse (2002: 205) points out that the bidimen-
sional approaches should be encouraged to motivate the strict delimitation
between these two theoretical notions. In cognitive approaches to aspect,
this bidimensional distinction is rarely encountered. Instead, perfective and
imperfective processes are conceptual entities that can be expressed by
other means than merely by morphological markings (cf. Langacker 1982).
Taking a functionalist vantage point of aspect, frequency information
takes a central role in grammatical description. While construals of predi-
cates display a remarkable versatility, they do not necessarily have the
same numerical distribution across various aspect constructions. Some con-
struals are only possible in certain contexts, some construals are very rare
and may even constitute hapax legomena in a corpus. In a usage-based
theory of language, investigating those verbs that are most strongly associ-
ated to the English Durative Adverbial Construction will reveal the default
semantics of this aspect construction. In addition, considering the construal
that associated verb stems receive in the verb-argument cluster as well as in
the construction will shed light not only on possible construals of predicates
in that construction, but also on the conventional construals.
The English durative adverbial belongs to the general cross-linguistic
class of atelic extend adverbials (cf. Haspelmath 1997). Atelic extend ad-
verbials are most often used with stative quantified situations (Haspelmath
1997: 39). While a broad and cross linguistic classification of constructions
serves well to capture commonalities in meaning and both form and distri-
bution of the specific construction, it does not capture the peculiar facts
about a specific construction in a particular language. Formally, the English
durative construction is unusual for an atelic extend adverbial, as these tend
to be marked cross-linguistically by either zero-expression or expression of
a minimal case (Haspelmath 1997: 120). The English durative construction,
140 Stefan Fuhs

however, is marked by the preposition for, and the temporal extend is most
often denoted by a noun phrase, expressing a canonical time period. Fol-
lowing Haspelmath (1997: 26) these can be grouped into three sub-types:
time units, calendar unit frames and qualitative periods. Of these sub-types,
only time units can occur in the Durative Adverbial Construction (Haspel-
math 1997: 26-7).
Semantically, the English Durative Construction has been described to
have an atelic meaning. For this reason it has been widely used as a test
item in the classical literature on aspect (c.f. Dowty 1979; Vendler 1967;
Verkuyl 1993; Haspelmath 1997; Smith 1997). The classical aspectual
classes of verbs – the Vendler classes – can in fact be distinguished by the
possible co-occurrence of the verbs with certain test frames, most notably
1
time adverbials. The treatment and the role it has received differ substan-
tially, however, since all major works written on the subject stem from
different theoretical paradigms. In the early work (Dowty 1979; Vendler
1967), the durative adverbial is introduced as a test frame for atelic predi-
1
cates . Traditional test frames as illustrated in (1) and (2) below would
check for ‘acceptability’ with the adverbial to determine the telicity values.
(1) He danced for an hour.
(2) *He died for an hour.
Both Vendler and Dowty are aware of exceptions to the rule, that is, those
cases in which a construal is actually possible, but only if the verb-
argument cluster receives a different construal than it would in the absence
of the Durative Adverbial Construction, as illustrated by the change in con-
strual in (3) and (4):
(3) The light flashed. [semelfactive reading]
(4) The light flashed for an hour. [iterative reading]
Coercion effects for aspectual constructions have been described in differ-
ent paradigms of linguistics. While there is agreement over some of the
basic directions of change, there is considerable disagreement over the
question of how this change comes about, what triggers it and which parts
carry the actual aspectual meaning.
A good overview of the possible directions of situation type shifts can
be found in Moens and Steedman (1988). They point out that aspectual
coercion, or type-shifting, is not a linear process that will ultimately come
to an end at which no possible construal is permissible anymore. Instead,
new linguistic material can result in new shifts and situation types can, in
 Aspectual coercion of the English Durative Adverbial 141

theory, be re-construed in endless loops (Moens and Steedman 1988: 21).

They provide the following possible (albeit artificially construed) example
to illustrate the ‘loopy’ nature of aspectual coercion:
(5) It took me two days to play the ‘Minute Waltz’ in less than sixty seconds
for more than an hour. (Moens and Steedman1988: 21)
The occurrence of aspectual coercion, however, is merely part of the pic-
ture. Another major issue in research on aspectuality is the nature of aspec-
tual composition. Aspectual types, as denoted by constructions and lex-
emes, tend to shift, depending on the elements present in the clause. The
sheer multitude of factors involved, including verbal morphology, lexemes,
grammatical properties of verbal arguments and more peripheral construc-
tions such as adverbials has led some scholars to conclude that the ways in
which aspectual factors interact is unknown. Other studies have attempted
to model the ways in which aspectual types can shift depending on the
presence and absence of relevant factors. Chang et al. (1998) provide a
dynamic model of aspectual composition that encodes goals and resources -
conceptual properties of aspectual situations. This model captures possible
paths of construal of the overall situation type in a computational network
by mapping out pathways of construal which can be triggered by the pres-
ence of linguistic constructions. To illustrate the difference, the sentences
in (6) and (7) display a minimal pair in this sense.
(6) He pushed the cart.
(7) He pushed the cart to the park.
In example (6), the construed situation is not telic. The activity of pushing
the cart does not necessarily lead to any particular endpoint, at least it is not
denoted by the predicate in this syntactic environment. By contrast, the
situation in (7) is telic, since there is a resource that is being consumed, i.e.
the distance between the subject referent and the park. The natural endpoint
of the situation is reached once the resource has been consumed entirely. A
further complication arises here since the aspectual factors (the resource)
do not map directly to the grammatical elements expressed in the clause
(the cart and the park). Neither is actually consumed by the action, the con-
strual operates on a verbal scale that only arises out of the presence of the
two, in other words, it is the distance between them that is sequentially
scanned until it is consumed entirely. This has been described as the incre-
mental theme (Hay, Kennedy & Levin 1999), or verbal scale (Croft
forthc.: 12).
142 Stefan Fuhs

In formal compositional approaches, coercion effects are explained in

terms of derived situation types. If, for instance, a basic situation type fails
at a feature check when combined with a temporal adverbial, then the basic
situation type will be re-written by a rule to abide by the requirements of
the temporal adverbial. The examples below illustrate such a derivation:
(8) Mary coughed.
(9) Mary coughed for an hour.
The first sentence in (8) is punctual. It denotes a singular occurrence of the
event, or a semelfactive (Smith 1997:53). This semelfactive feature set
clashes with the requirements of the durative adverbial in (9) and conse-
quently, the features of Mary coughed have to be re-written to be durative
themselves. The outcome is the iterative reading obtained in (9). Expressed
in the notation applied by this approach, the rule of derivation can be writ-
ten down as such:
(10) Mary coughed for an hour.
VCon[+Dyn -Telic -Dur] + Adv[+Dur] -> DVCon[+Dyn -Telic +Dur]
Coercion effects in sign-based treatments of aspectuality capture similar
phenomena, though the role of the construction in aspectual composition is
radically different. The aspect construction in a sign-based approach de-
notes aspectual types in the same manner that predicates do. Here, aspec-
tual coercion is the consolidation of constructional requirements and predi-
cate meaning. As such, there is no inherent difference between
constructional concord such as (11) and constructional type-shifting, or
coercion, as in (12)
(11) He read a book in an hour.
(12) He died in an hour.
In both cases, the semantics of the aspect construction require a phase
change as part of the profile. In the case of (11), read a book is available as
an accomplishment construal, so the profile of the predicate already in-
cludes duration and the phase change. In (12) a profile-shift is required to
make the otherwise punctual achievement predicate [DIE] available to such
a construal; the profile is shifted to highlight the duration leading up to the
phase change of the aspectual contour. This case represents a runup
achievement, following Croft (forthc.: 15). While one case illustrates clas-
sical concord and the other one classical coercion, the mechanics are quite
the same. The construction’s aspectual denotation requires the predicate to
 Aspectual coercion of the English Durative Adverbial 143

be available as a construal and as long as such a construal is available, con-

structional and predicate denotation can be unified successfully, i.e. the
situation is interpretable.

3. Methods

3.1. Collostructional Analyses

The Collostructional Analysis is one of the more established methods of

corpus linguistic investigation. It has been applied to a range of empirical
studies from different areas of linguistic description and it has been well-
documented and illustrated in a collection of methodological papers (cf.
Stefanowitsch and Gries 2003, Gries and Stefanowitsch 2004, Ste-
fanowitsch and Gries 2005). It is theoretically based on a sign-based view
of language structure, in which form-meaning pairings, or constructions,
are present on all levels of grammatical description. In other words, lex-
emes, morphemes, phrases and more complex syntactical patterns differ in
their schematicity and complexity, but not in their fundamental makeup.
They all are form-meaning pairings and they all denote. In such a frame-
work, the semantic meaning of a grammatical construction can be described
by investigating the lexemes that co-occur in a prominent slot in that con-
struction – the collexemes. While simple frequency lists of lexemes can
illustrate what semantic material is possible in that particular slot, even
whether it is frequent or infrequent, it cannot tell us whether that particular
lexeme occurs more often in that construction than random distribution
would suggest. Lexemes that are highly frequent in a language may also
occur frequently in certain constructions, but since they also occur fre-
quently elsewhere, they are not particularly associated to that construction.
Association, however, is fundamental to grammatical description. This
circumstance has been incorporated into linguistic models, such as usage-
based models (cf. Barlow and Kemmer 2000; Croft and Cruse 2004) and
support for the psychological importance of such distributional findings has
come from recent studies on converging evidence from corpus linguistics
and psycho-linguistic experiments (cf. Gries, Hampe and Schönefeld 2005).
The Collostructional Analysis offers an empirical method to identify
and rank collexemes not merely by their raw frequency, but by their asso-
ciation to the investigated construction. The output of this method allows
inferences about the semantics of the construction by identifying collexeme
144 Stefan Fuhs

classes among the significantly attracted collexemes. In order to calculate

the association strength of each collexeme, several pieces of information
have to be collected by corpus-linguistic means: (i) the frequency of the
collexeme in the construction, (ii) the frequency of the collexeme in all
other constructions in the corpus, (iii) the frequency of other lexemes in the
construction in the corpus and (iv) the frequency of all other lexemes in
other constructions in the corpus (Stefanowitsch and Gries 2003: 218).
These data are cross-tabulated in a 4x4 table for each lexeme that occurs in
the construction. To calculate the association strength, several association
measures are possible, but for all practical purposes the Fisher-Exact test is
taken as a default, both for its ability to retain precision even with small
sample sizes and for fitting the distributional reality of linguistic data (cf.
Stefanowitsch and Gries 2003; Pederson 1996).
To obtain the information for this 4x4 table, each Collostructional An-
alysis starts by retrieving all instances of the construction from the corpus.
Collexemes are coded and counted and the remaining table cells are ob-
tained by performing additional corpus queries. The actual calculation of
the Fisher Exact test is complex and poses a heavy computational load, but
contemporary computer hardware is more than equipped to handle is.
Therefore, the calculations are performed by software. In this case, an R
script (Gries 2007) was used to run the Collostructional Analysis once the
raw frequency tables had been obtained, as elaborated below.

3.2. Data and coding

As a first step to investigating the constructional semantics of the Durative

Adverbial Construction, all instances of the durative adverbial had to be
extracted manually from a corpus. The said construction is formally ex-
pressed by a prepositional phrase which is headed by for, followed by a
string denoting a time interval, i.e. [for TimeInterval]. This time interval
can be expressed by a simple noun phrase, as in for three weeks, but it can
also be more complex, for instance by being represented by an as-
predicative such as for as long as it takes. Since the time interval can be
represented by such varied forms, it is hard to define a search pattern that
will not exclude correct tokens of the construction (false negatives) or in-
clude strings that do not represent the construction (false positives). Creat-
ing a query for for alone would yield a massive amount of false positives if
no other factors are included in the search. For this reason, the Interna-
 Aspectual coercion of the English Durative Adverbial 145

tional Corpus of English – British component (ICE-GB) was selected as the

corpus for this investigation. The advantage it offers over much larger cor-
pora, such as the British National Corpus (BNC hereafter), is that it comes
with functional tagging and more to the point, it allows the user to search
for adverbials which are headed by for, which greatly delimits the amount
of false positives that a comparable BNC query would unearth.
To extract all instances of the durative construction from the corpus in a
way that would not falsely exclude any correct tokens, I queried for all
adverbials headed by the preposition for. This initial search yielded 6232
tokens which were then manually inspected. Even at a fleeting glance, the
set still included many false positives, the majority of which were preposi-
tional datives. I also excluded all tokens of formally related constructions
which nonetheless have distinct aspectual semantics, most noticeably the
[for the nth time] construction, as illustrated by (13):
(13) Mrs Mandela sat impassively while Mr Kgase gave evidence in the court,
which was half-empty for the first time since the case opened last month.
(ICE-GB:W2C-019 #066:4)
This aspect construction requires an entirely different construal of the
predicate than the durative adverbial and hence has to be excluded. Clauses
containing further adverbials were also dismissed, since their influence on
the predicate construal could not be determined and hence they had to be
controlled for. In addition, cases of negation and cases containing verbs of
negative polarity were also removed manually, since the aspectual types
represented by such clauses were often impossible to determine using the
ready-made tests.
In coding, I followed Smith’s suggestion to classify single for-phrases
as instances of the durative construction, whereas in double marking, the
second phrase was coded as a ‘purpositive extend’ (Haspelmath 1997: 128;
Smith 1997: 114). After manually clearing the set, a total of 671 true in-
stances of the English Durative Adverbial remained, which occurred in
clauses in which no other aspect constructions in the periphery were present
that could influence the predicate construal. This set was then used for a
Collostructional Analysis.
146 Stefan Fuhs

4. Results and discussion

Let us first investigate the distribution of aspectual types, denoted by the

V+ARG clusters with which the Durative Adverbial Construction com-
bined. The prediction in this case is that the Durative Adverbial ‘selects’
atelic predicates. While coercion effects have certainly been described, they
have still received marginal attention in the description of the construction.
As presented in table 1 below, the actual distribution of aspectual types
does indeed show a preference for atelic construals: approximately 75% of
all tokens were indeed either states (14) or activities (15). However, a sur-
prising 25% of all tokens were actually achievements (16), in which the
telic predicate was reconstrued to be available in the durative’s atelic
denotion. In (16), the telic and punctual predicate fall asleep is made avail-
able as a possible construal by a profile shift. Instead of profiling the punc-
tual and telic inception of the event, the predicate combined with the adver-
bial in (16) now profiles the result state obtained after the inception (cf.
2
Croft forthc. ).
(14) And they’re apart just for a short period of time. (ICE-GB:S2A-056
#130:2:A)
(15) Yeah but they’ve been saying that for ten years. (ICE-GB:S1A-029
#177:1:B)
(16) Well I fell asleep for two hours. (ICE-GB:S1A-011 #101:1:C)
The results of the Collostructional Analysis are presented in table 2 below.
This three-part list ranks all significantly attracted collexemes in predicates
combined with the durative adverbial in their order of association to that
adverbial. The list itself is ranked from top to bottom – highly associated
collexemes are at the top, less strongly-associated collexemes at the bot-
tom. The association strength is presented in the form of collostructional
strength. To convert this to a p-value, the following conversion can be ad-
opted: coll.strength>1.30103 => p<0.05 (Gries 2007). That is to say, every-
thing on the list is significantly attracted, and the higher the association
strength, the more significant the association.
At a first glance, the collexeme list in Table 1 provides an expected pic-
ture. The most strongly-associated collexemes of the Durative Adverbial
Construction are stay, remain and last. These verb types are stative (assum-
ing an out-of-context interpretation) and as such they are atelic. Next on the
list and still also highly significantly associated are chill, process and work,
three verb types that can denote activities. As such they are dynamic and
 Aspectual coercion of the English Durative Adverbial 147

atelic. The next verb type, by contrast, seems to be unexptected in this con-
struction. Go away, significantly associated with a collostructional strength
of ~4.643, is an achievement and hence telic. Another odd-man-out is pres-
ent in the form of stop, an aspectual verb or terminative aspect construction
that denotes a punctual phase-change, i.e. an achievement. This, too, is
telic. More of these telic verb types follow: put.out.of (coll.str=2.3138),
slip.away (coll.str=1.8388), shut.down (coll.str=1.8388) and fall.asleep
(coll.str=1.4181).
Table 1. Significantly attracted collexemes of the English durative construction
verb coll.strength verb coll.strength verb coll.strength
stay 21,60713164 leave.free 2,313804284 muse 1,714903676
remain 8,038168769 smooch 2,313804284 eye 1,714903676
last 7,720215453 put.out.of 2,313804284 gush 1,714903676
chill 6,943350435 fret 2,313804284 prop 1,619045092
process 6,794051333 gas 2,313804284 grip 1,619045092
work 4,841131913 harrass 2,013828266 speak 1,588513385
go.away 4,643087621 enthuse 2,013828266 put.off 1,540914425
simmer 4,628254105 batter 2,013828266 grill 1,540914425
delay 4,402597508 elude 2,013828266 exist 1,512822397
go 4,199299257 talk 1,9378052 wear 1,512822397
dwell 3,189510313 wait 1,916156778 be 1,491898988
sleep 3,038993428 slip.away 1,838790135 tolerate 1,475017365
keep 2,986442273 carry.around 1,838790135 sit 1,420646223
stop 2,825088554 cover.for 1,838790135 fall.asleep 1,418074298
lie 2,688697433 shut.down 1,838790135 reelected 1,418074298
drag 2,515712748 wizz 1,838790135 admire 1,418074298
live 2,472269382 digress 1,838790135 fry 1,367969805
rent 2,465955543 distract 1,838790135 enrich 1,367969805
continue 2,425234034 reign 1,838790135 argue 1,346907764
go.beserk 2,313804284 develop 1,838790135 visit 1,339496314
toil.away 2,313804284 retain 1,735015932 hang 1,339496314

These results are noteworthy for two reasons: (i) telic predicates can in fact
be combined with the atelic durative adverbial – as apparent from the dis-
tribution in Table 1 and, moreover, some of these telic predicates are in fact
significantly attracted to this atelic construction. They do not produce a
‘clash’ with the constructional semantics, but instead they are available as a
construal in the Durative Adverbial Construction via a profile-shifting op-
eration. Interestingly, it is this construal for this particular verb type that is
in fact associated with the construction. In all cases in which a telic predi-
cate occurred in the construction, it was in the form of an achievement,
148 Stefan Fuhs

which was then reconstrued as denoting the resultant state after the phase
change.
So far, the association measures presented in table 2 investigate the oc-
currence of two forms: the construction and the verb stem. Without further
investigation, this association can be misleading if taken at face value. The
telic verb stem could already have been re-construed by formal elements in
the predicate, i.e. the progressive construction. More to the point, a telic
verb, occurring in the imperfectivizing progressive, does not have a telic
construal anymore. It is, in fact, part of an atelic predicate. In such a case,
the co-occurrence with the durative adverbial would really have to be at-
tributed to the fact that the entire predicate is atelic and hence available as a
construal in the durative adverbial without re-construing it via profile-
shifting. The Collostructional Analysis as such would be ‘blind’ to such
intermediate construal steps and manual inspection of the original data is
required. The examples in (17) and (18) below illustrate such a case:
(17) Shells exploded in the outskirts of Sarajevo.
(18) Shells were exploding in the outskirts of Sarajevo.
A closer look at the tokens of the telic verb stems listed above was con-
ducted to address this point. Every occurrence in the original data was
checked for co-occurrence of the progressive in that clause, the results are
presented in Table 2 below:
Table 2. Co-occurence of telic collexemes and the progressive
collexeme aspect.PROG aspect.PLAIN
go.away + +
stop - +
go.berserk - +
toil.away - +
put.out.of - +
slip.away - +
shut.down - +
fall.asleep - +

The results of this inspection demonstrate that in fact only for one of the
collexemes under investigation, the progressive construction entered the
scene as a possible source of atelic construal. For all others, the telic verb
stems did in fact receive a telic interpretation in a telic predicate, i.e. no
progressive construction was present. The significant association discussed
 Aspectual coercion of the English Durative Adverbial 149

above is therefore not epiphenomenal in the sense that it neglects the influ-
ence of an important aspect construction. Instead, it is a direct and valid
association between a telic verb stem in a telic predicate construal and the
English Durative Adverbial Construction.
This is further illustrated by a complementary Collostructional Analysis
of a subset of the English durative construction. While the first analysis
included all tokens of the construction, those with achievement construals
as well as those with activitiy or state construals, the next two analyses
broke this set up into telic construals (achievements) and atelic construals
(activities, states). The collexemes of the Durative Adverbial Construction
combining with telic predicates are presented in Table 3 below:
Table 3. Significantly attracted collexemes of the English Durative Adverbial Con-
struction with a telic core construal (achievements).
verb coll.strength verb coll.strength
go.away 6.1915183857882 shut.down 1.84080233236444
chill 4.15657409851742 slip.away 1.84080233236444
toil.away 2.31582627695936 digress 1.84080233236444
go.beserk 2.31582627695936 lie 1.79161669406308
stop 2.02361018179313 prop 1.62104752596006
enthuse 2.01584535721543 grill 1.54291198889905
simmer 2.01584535721543 reelected 1.42006214498845
elude 2.01584535721543 dwell 1.42006214498845
put.out.of 1.84080233236444 fall.asleep 1.42006214498845
distract 1.84080233236444 fry 1.36995280624343

The collexeme list above shows verbs and verb-particle constructions that
are usually interpreted in telic construals. The first set includes achieve-
ments such as go.away (coll.str=6.19), go.berserk (coll.str=2.32), put.out.of
(coll.str=1.84) and shut.down (coll.str=1.84). These verbs were part of
predicates in which they had a telic construal. This telic predicate in turn
was available in the durative via profile-shifting, from the actual punctual
phase change to the resulting state. In other words, the phase change in (19)
is still part of the aspectual contour, but the whole unified construction
profiles the state of being in Dorset, instead of the act of going there. In a
similar manner, the phase change denoted by put.out.of.action in (20) is no
longer profiled by the combined structure of predication and adverbial con-
struction. Instead, the resulting state is profiled. In both cases, the resulting
state is bounded again by the adverbial – the termination is also part of the
contour, but not part of the profile.
150 Stefan Fuhs

(19) Well we went away to Dorset for a week at Easter. (ICE-GB:S1A-021

#250:2:C)
(20) I got into this because cerebro-spinal meningitis put me out of action for
six months in nineteen fifty-two <,> and as few barristers’ practices can
stand that sort of interruption I applied for a job,(…)( ICE-GB:S2B-025
#003:1:A)

Table 4. Significantly attracted collexemes of the English Durative Adverbial Con-

struction with an atelic core construal (states, activities)
verb coll.strength verb coll.strength
stay 18.8010482732977 retain 1.94627493937510
remain 8.85971185947352 wizz 1.91834402731464
last 8.26582785917278 carry.around 1.91834402731464
process 7.18866456196719 cover.for 1.91834402731464
work 5.0347577581907 chill 1.91834402731464
see 4.93868940718354 reign 1.91834402731464
delay 4.71309336102246 develop 1.91834402731464
be 4.57117765769294 muse 1.7942805256178
live 2.91730876092689 eye 1.7942805256178
continue 2.71075127981652 gush 1.7942805256178
drag 2.67212236610695 exist 1.71780154252650
rent 2.62213087309853 grip 1.69824516098926
smooch 2.39371305265399 put.off 1.61993797678082
leave.free 2.39371305265399 tolerate 1.55386466244163
fret 2.39371305265399 admire 1.49674560420116
gas 2.39371305265399 dwell 1.49674560420116
wait 2.18758907975370 hang 1.48373637872592
sustain 2.15442717293559 enrich 1.44646538393130
keep 2.15385071578597 hold 1.42504133747857
harrass 2.09355946493628 witness 1.32414003445723
batter 2.09355946493628 speak 1.30326156496843
simmer 2.09355946493628
go.on 2.07738584360067
sleep 1.96129531165755

For comparison, a subset of the construction was considered, in which all

verbs are part of verb-argument clusters in which they received an atelic
construal. The results for a Collostructional Analysis performed on these
verb-argument clusters with an atelic construal are presented in Table 4.
Worth noting, the verbs attracted to the durative adverbial as part of atelic
predicates are atelic in their default interpretations. The most strongly asso-
ciated collexemes stay (coll.str=18.8), remain (coll.str=8.81) and last
 Aspectual coercion of the English Durative Adverbial 151

(coll.str=8.26) are the ones strongly attracted to the construction. These

‘atelic’ verb stems are not involved in a profile-shift. Instead, the phase
profiled by the predicate is measured out by the durative adverbial. This is,
put differently, the function of the adverbial that is most often evoked in
aspectual discussion and telicity testing.

5. Conclusion

The study has shown that the English durative construction is not limited to
the co-occurrence with atelic predicates. Coercion effects, in which the
atelic construction yields a re-construal of an otherwise telic predicate,
make up as much as 25 per cent of all actual tokens and hence the theoreti-
cal ‘odd-man-out’ is more frequent than current qualitative treatments may
at least suggest.
Furthermore, the Collostructional Analysis showed a significant attrac-
tion of the construction to verb stems that are held to be telic in their default
construals. These verb stems were embedded in a syntactical environment
in which they had not undergone a re-construal in which they would have
denoted an atelic situation type and hence there is indeed a significant at-
traction between the construction and a specific construal of a telic verb
stem.
In a broader perspective, this study, which merely sheds a spotlight on
one particular aspect construction, points towards a more dynamic interplay
between verb semantics and constructional semantics in aspectual compo-
sition. If a construction such as the durative adverbial can be so directly
involved in the aspectual makeup of clauses, then the picture of aspec-
tuality must include all elements of the clause, not merely verbs and their
arguments in determining the aspectual type denoted by that entire clause.
Constructions situated in the periphery of the clause are part of aspect just
as verb stems are and to address Sasse’s initial question: adverbials are part
of the aspectual makeup of clauses and their presence does trigger re-
construals of certain predicates.
In closing, I’d like to stress that this brief investigation of the interaction
of grammatical constructions and verbs merely offers a first glimpse into
the interaction of the two in aspectual makeup, but that it points us to a very
fundamental insight – aspectual constructions both denote and influence
aspectuality in the same way that verbs are known to do and their constru-
als are closely tied. Ultimately, this challenges the still prevalent idea that
152 Stefan Fuhs

aspectual types, a conceptual category, is directly tied to the verbs’ seman-

tics, i.e. that we have reason to treat verbs as if they were activity verbs,
achievements verbs and such. Instead, the construal of verbs seems to be
much more lenient than this strict delineation into aspectual verb classes
suggest and the role of the durative adverbial in evoking certain construals
in selected verbs should be emphasized instead.

Notes

1. Vendler (1967) uses the term term to describe both verbs and verb+argument
clusters. It is debatable whether the original ‘Vendler classes’ of verbs really
were corresponding to verb types proper or whether the author included as-
pectual influence of the arguments. In most contemporary theories, aspectual
classes are assigned to VPs, rather than verb stems.
2. To be precise, achievements that do not profile a result state after their incep-
tion are unavailable in this construction. Only directed achievements and cy-
clic achievements are available (Croft forthc.: 26).

References

Barlow, Michael and Suzanne Kemmer (eds.)

2000 Usage-Based Models of Language. Stanford: CSLI Publications.
Chang, Nancy, Daniel Gildea and Srini Narayanan
1998 A dynamic model of aspectual composition. In Proceedings of the
20th Cognitive Science Society Conference. Morton Ann
Gernsbacher and Sharon J. Derry (eds.), 226-231. Mahwah, N.J.:
Lawrence Erlbaum.
Comrie, Bernard
1976 Aspect. Cambridge: Cambridge University Press.
Croft, William
1999 The structure of events and the structure of language. In The New
Psychology of Language, Michael Tomasello (ed.), 67-92. Mahwah,
N.J.: Lawrence Erlbaum.
forthc. Verbs: Aspect and argument structure. Oxford: Oxford University
Press.
 Aspectual coercion of the English Durative Adverbial 153

Croft, William and D.Alan Cruse

2004 Cognitive Linguistics. Cambridge: Cambridge University Press.
Dowty, David R.
1979 Word Meaning and Montague Grammar. Dordrecht: Reidel.
Gries, Stefan Th. and Anatol Stefanowitsch
2004 Extending collostructional analysis: A corpus-based perspective on
‘alternations’. International Journal of Corpus Linguistics 9(1): 97-
129.
Gries, Stefan Th., Beate Hampe and Doris Schönefeld
2005 Converging evidence: Bringing together experimental and corpus
data on the association of verbs and constructions. Cognitive Lin-
guistics 16(4): 635-76.
Gries, Stefan Th.
2007 Coll.analysis 3.2. A program for R for Windows 2.x.
Haspelmath, Martin
1997 From Space to Time: Temporal adverbials in the world’s languages.
Munich: Lincom Europa
Hay, Jennifer, Christopher Kennedy and Beth Levin
1999 Scalar structure underlies telicity in ‘degree achievements’. In The
Proceedings from the Ninth Conference on Semantics and Linguistic
Theory: SALT IX. Ithaca, New York. Cornell University, Depart-
ment of Linguistics. CLC Publications.
Langacker, Ronald W.
1982 Remarks on English aspect. In Tense-Aspect: Between semantics &
pragmatics, Paul Hopper (ed.), 265-304. Amsterdam / Philadelphia:
John Benjamins.
Michaelis, Laura A.
2004 Type shifting in Construction Grammar: An integrated approach to
aspectual coercion. Cognitive Linguistics 15: 1-67.
2005 Entity and event coercion in a symbolic theory of syntax. In Con-
struction Grammars: Cognitive grounding and theoretical exten-
sions, Jan-Ola Östmann and Miriam Fried (eds.), 45-88. Amsterdam
/ Philadelphia: John Benjamins.
Moens, Marc and Marc Steedman
1988 Temporal ontology and temporal reference. Computational Linguist-
ics 14(2): 15-28.
Pedersen, Ted
1996 Fishing for exactness. In Proceedings of the South-Central SAS
Users Group Conference, Austin, TX, 188–200.
154 Stefan Fuhs

Sasse, Hans-Jürgen
2002 Recent activity in the theory of aspect: Accomplishments, achieve-
ments, or just non-progressive state? Linguistic Typology 6: 199-
271.
Smith, Carlotta
1997 The Parameter of Aspect (2nd ed.). Dordrecht: Kluwer.
Stefanowitsch, Anatol and Stefan Th. Gries.
2003 Collostructions: investigating the interaction between words and
constructions. International Journal of Corpus Linguistics 8(2): 209-
43.
Stefanowitsch, Anatol and Stefan Th. Gries
2005 Covarying collexemes. Corpus Linguistics and Linguistic Theory
1(1): 1-43.
Vendler, Zeno
1967 Linguistics in Philosophy. Ithaca: Cornell University Press.
Verkuyl, Henk J.
1993 A Theory of Aspectuality: The interaction between temporal and
atemporal structure. Cambridge: Cambridge University Press.
The force dynamics of English complement clauses:
A Collostructional Analysis

Martin Hilpert

Abstract
This paper studies English gerund clauses with an infinitive complement, such as
Learning to read is fun, through a quantitative corpus analysis of co-occurring
lexical material. Previous cognitive linguistic studies have presented general ac-
counts of gerund complement clauses, but have not focused on this particular con-
struction. Adopting the theoretical framework of Construction Grammar, it is ar-
gued here that the construction is a symbolic unit of English grammar that is
endowed with force dynamic meaning. The meaning of force dynamics, which
encompasses for example the resistance to a force, the use of force to overcome an
obstacle, or the balance of two counteracting forces, pervades grammatical do-
mains such as modality and causation; the present paper makes the case that it
extends also to the domain of complement clauses. It is further illustrated that a
quantitative, corpus-linguistic approach can not only yield a detailed semantic
description of a grammatical construction, but that it can also make predictions that
adequately mirror the intuitions of native speakers.
Keywords: force dynamics, gerund complement clauses, Collexeme Analysis,
nominalization, factive, action-referrent

1. Introduction

This paper presents a quantitative corpus-based analysis of English gerund

clauses with an infinitive complement, such as Learning to read is fun.
Previous cognitive linguistic studies have presented general accounts of
gerund complement clauses (Langacker 1991, Heyvaert 2003), but have not
focused on this particular construction. Adopting the theoretical framework
of Construction Grammar (Goldberg 2006), this paper argues that this con-
struction type is a symbolic unit of English grammar that is endowed with
meaning and that needs to be described in its own right.
The present study of gerund clauses is usage-based and draws on data
from the British National Corpus. A quantitative analysis of co-occurring
156 Martin Hilpert

lexical material (Stefanowitsch and Gries 2003) is used to investigate the

preferences that the construction exhibits for particular combinations of
gerunds and infinitive complements. The analysis shows that the construc-
tion is associated with lexical material that can be characterized in terms of
force dynamics (Talmy 2000), which encompasses for example the resist-
ance to a force, the use of force to overcome an obstacle, or the balance of
two counteracting forces. The meaning of force dynamics pervades gram-
matical domains such as modality and causation; it is argued here that it
also lies at the heart of the investigated complement clause construction.
The remainder of this paper is structured as follows. Section two sur-
veys previous studies of gerund clauses, both from the cognitive tradition
as well as from formal approaches, that make relevant predictions for the
present analysis. The third section introduces central concepts of force dy-
namics, such as the agonist and the antagonist, which will be of importance
in the semantic description of the investigated construction. Section four
presents the methodology used in this paper, explaining how Collexeme
Analysis (Stefanowitsch and Gries 2003) is applied to the construction
under investigation. In section five, the results of the analysis are integrated
into a semantic description of the construction that explains how the no-
tions of force dynamics pertain to the constructional meaning. Section six
takes a step back and discusses the results from a broader theoretical per-
spective. It is concluded that a quantitative, corpus-linguistic approach can
not only yield a detailed semantic description of a grammatical construc-
tion, but that it can also make predictions that adequately mirror the intu-
itions of native speakers.

2. Previous work

This section gives a brief definition of the investigated construction and

addresses previous findings about gerund clauses that are relevant for the
present study. Gerund complement clauses are generally recognized as a
nominalization strategy in English (Quirk et al. 1985). Several formal
treatments of gerundive nominalization (Lees 1960, Schachter 1976,
amongst others) have offered syntactic and semantic analyses of sentences
such as Going to the beach is fun or Having attended Harvard can be help-
ful. This section reviews some of their observations, and also discusses the
functional accounts of English nominalization in Langacker (1991) and
Heyvaert (2003).
 Force dynamics of English complement clauses 157

Gerund complement clauses are gerundive phrases that assume the func-
tion of the grammatical relations subject and object. While there are a num-
ber of different types of gerund complement clauses, this study focuses on
gerund clauses with a second, non-finite verb that forms a syntactic con-
stituent together with the gerund. In the following examples, the gerund
phrase in (1a) is a subject complement clause whereas in (1b) it is an object
complement clause.
(1) a. Learning to read is an enjoyable experience for children.
b. Children enjoy learning to read.
Regardless of which grammatical relation is instantiated by the gerund
clause, the morphosyntactic form of the construction always comprises a
gerund, the infinitive marker to, and a non-finite verb. However, specifying
this form does not yield a complete definition of the construction. Colloca-
tions such as learning to read do not necessarily instantiate a gerund clause,
as is illustrated in (1c).
(1) c. When children are learning to read, they usually enjoy the experience.
In this example, learning to read does not function as a grammatical rela-
tion, but instead forms part of a verb phrase. A definition of the construc-
tion under investigation must therefore make reference to both the morpho-
syntactic form V-ing to V and the larger syntactic context.
Gerund complement clauses are known to have typical semantic charac-
teristics. A fundamental distinction is pointed out by Lees (1960: 58), who
distinguishes factive and action-referrent gerund clauses. Factive gerund
clauses denote a particular state of affairs, whereas action-referrent gerund
clauses denote a process. Both meanings can be expressed by the investi-
gated construction. A factive example is given in (2a); an action-referrent
example is shown in (2b).
(2) a. Having to leave is the saddest thing.
b. Planning to go on a holiday is fun.
Lees argues that the semantic difference between the two types is reflected
in a syntactic characteristic. Factive gerund clauses can appear with geni-
tive pronouns, which action-referrent gerund clauses cannot.
(2) c. His having to leave is the saddest thing.
d. *His planning to go on a holiday is fun.
Schachter (1976: 215) aims to explain this syntactic difference by pointing
out that genitive pronouns have the function of making their complements
158 Martin Hilpert

definite, that is, tied to a specific situation. In example (2a), bare having to
leave evokes a generic situation, whereas his having to leave in (2c) refers
to a definite instance of such a state. Schachter concludes that only factive
gerund clauses can be made definite. The question that remains is why facts
such as having to leave can successfully be projected onto definite situa-
tions, while activities such as planning to go on a holiday cannot.
Langacker (1991: 32) takes issue with the discussed observations. First,
he argues that the distinction between factive and action-referrent gerund
clauses cannot be made on the ground of the lexical semantics of the ge-
rund alone. It is a matter of construal that in (3a), taunting is interpreted as
a process, whereas in (3b) it is understood as a fact.
(3) a. Harvey’s taunting of the bear was merciless. (a-b: Langacker 1991: 32)
b. Harvey’s taunting of the bear came as a big surprise.
Langacker’s observation carries over to the investigated construction, as is
illustrated in (4a) and (4b). The complement clause learning to read can be
understood as either denoting a process or a fact.
(4) a. Learning to read is an enjoyable experience.
b. Learning to read was the last thing we expected from Harvey.
A second point of Langacker’s concerns Lees’ (1960) claim that action-
referrent gerund clauses cannot occur with genitive pronouns. Example (3a)
shows that this generalization is not always true, as it occurs with a genitive
nominal. The grammatical properties of factive and action-referrent gerund
clauses thus need to be reconsidered. Langacker proposes that action-
referrent gerund clauses behave more like nominals, whereas factive ge-
rund clauses share behavioral properties of finite clauses. For example,
action-referrent gerund clauses may take an attributive adjective (5a) or a
definite determiner (5b), whereas factive gerund clauses may occur with an
adverb (5c), and usually do not allow a determiner (5d).
(5) a. Zelda’s reluctant signing of the contract. (a-d: Langacker 1991: 31)
b. the signing of the contract.
c. Zelda’s reluctantly signing the contract.
d. *the signing the contract.
With respect to the present analysis, we can take Langacker’s observations
to predict that the investigated construction should exhibit a bias towards
action-referrent meaning. This follows from the fact that complement
clauses, in their role as subjects or objects, function as nominals rather than
as finite clauses.
 Force dynamics of English complement clauses 159

Heyvaert (2003) follows Langacker’s suggestion that factive gerund

clauses exhibit clausal behavior while action-referrent gerund clauses dis-
play nominal properties. However, while Langacker argues that gerund
clauses denote undetermined instances of processes, Heyvaert views them
as denoting atemporalized process types (2003: 222). While this appears a
very subtle distinction, it actually makes a testable prediction for the pres-
ent analysis. If Heyvaert is correct in assuming that the investigated con-
struction denotes atemporal types rather than instances, this should be re-
flected in the lexical material that occurs in and around the construction in
actual corpus data. Examples (6a) and (6b), repeated here for convenience,
serve to illustrate this point.
(6) a. Learning to read is an enjoyable experience.
b. Learning to read was the last thing we expected from Harvey.
Whereas (6a) makes a fully generic statement that presents learning to read
as an atemporalized and decontextualized situation type, example (6b) ties
this process to a particular instance. On Heyvaert’s account, sentences like
(6a) should be found more frequently, because the overall generic meaning
of the sentence harmonizes with the generic meaning of the complement
clause. In (6b), by contrast, there is a mismatch between the generic com-
plement clause and the specific case of Harvey, so that examples like this
one should be dispreferred.
To summarize previous findings, the gerund complement clause con-
struction with two lexical verbs cuts across the distinction of factive and
action-referrent gerund clauses. Whether a particular example is factive or
action-referrent does not follow directly from the lexical semantics of the
involved verbs, but depends on its construal. With Langacker (1991: 31),
we hypothesize the investigated construction to express action-referrent
meaning more often than factive meaning. With Heyvaert (2003: 222), we
hypothesize that the construction expresses generic situation types more
often than specified instances. These hypotheses will be taken up again in
section four, which discusses how they can be tested against corpus data.

3. Force dynamics

Talmy (2000: 409) argues that the semantics of many grammatical domains
and many lexical items is force dynamic, such that their meanings reflect
‘the exertion of force, resistance to such a force, the overcoming of such a
160 Martin Hilpert

resistance, blockage of the expression of force, removal of such blockage,

and the like’. Such meanings underlie modal verbs like must, but also con-
junctions such as because, or prepositions like against. Examples (7a) to
(7c) give some illustrations.
(7) a. You must wear overalls when using large power tools.
b. People were evacuated because the area was flooded.
c. The rain hit against the window.
The common trait that underlies the meanings of these force-dynamic ele-
ments is the opposition of two entities with inherent forces. These two enti-
ties are perceived as figure and ground (Rubin 1958) of the same basic
scene. In example (7a), the ground is instantiated by an authority. This
authority exerts a force on the addressee You, who is the more prominent
entity and thus instantiates the figure. Talmy abandons the traditional ter-
minology of figure and ground, calling the more prominent force entity the
agonist and the opposing entity the antagonist.
Both the agonist and the antagonist carry a certain force which may
have either a tendency towards action or a tendency towards rest (Talmy
2000: 414). In example (7a), the agonist has a tendency towards rest, that
is, no inherent inclination to wear overalls. The antagonist aims to over-
come this tendency and to make the agonist act. To illustrate this further,
the tendencies towards rest and action are reversed in example (7d). Here,
the agonist has a natural tendency to eat ice cream that the antagonist does
not need to enforce.
(7) d. You may eat an ice cream when you have finished your veggies.
An important part of the constellation of forces between agonist and an-
tagonist is the strength of the respective forces (Talmy 2000: 414). In ex-
ample (8a), the force of the agonist is weaker than the force of the antago-
nist, while (8b) shows the reverse constellation. Depending on the balance
of strengths, the result of the force interaction can be either a state of rest or
a state of action.
(8) a. I couldn’t push the door open.
b. I managed to push the door open.
The constellation of an agonist and an antagonist with different inherent
tendencies and forces of varying strength allows for an array of different
force-dynamic meanings, such as causing, letting, stopping, preventing, and
many others. As Talmy points out, many of these meanings correspond to
grammaticalized forms such as causative constructions, modal verbs, and
 Force dynamics of English complement clauses 161

prepositions, in English and in other languages. In section five of this pa-

per, it will be argued that gerund clauses with a to-infinitive instantiate
precisely such a grammaticalized expression of force dynamics.

4. Theory and methodology

While the present analysis draws on the insights of the works discussed in
section two, it also departs from them in several aspects. First, it focuses on
a subtype of what previous analyses have viewed as basically one construc-
tion. Instead of approaching the construction family of gerund clauses as a
whole, we restrict our view to examples like Trying to find a job is quite
challenging. Second, the present analysis does not concern itself with ana-
lyses of underlying structures (Lees 1960, Pullum 1991) or syntactic char-
acteristics that gerundive nominals have in common with other construc-
tions of English (Schachter 1976, Langacker 1991). Lastly, while the
mentioned studies have been based on introspective data, this study offers a
corpus-based account of gerund complement clauses. This approach allows
us not only to put the investigation on an empirical basis, it also allows us
to make observations about the construction that can only be made through
quantitative data. This concerns especially the mutual attraction of the con-
struction with particular lexical material. As will be explained in more de-
tail below, the semantic interplay between lexical elements in the construc-
tion allows for a data-driven description of the constructional semantics.
The present study employs the quantitative method of Collexeme An-
alysis (Stefanowitsch and Gries 2003), which is a member of the family of
collostructional methods. In recent years, the papers of its two developers
have inspired a number of other researchers to experiment with the methods
and use them for different purposes (Collemann 2009, Gilquin 2006, Hil-
pert 2008, Speelman, Tummers and Geeraerts 2009, Wiechmann 2008,
Zeschel 2008, see also the contributions in this vol.). The method of Col-
lexeme Analysis, which is applied here, is a corpus-based analysis of collo-
cations which hinges on the assumption that grammar is a large inventory
of symbolic units, that is, constructions (Goldberg 1995, 2006). To make
this assumption more explicit, we adopt the following definition of con-
structions (Goldberg 2006: 5):
Any linguistic pattern is recognized as a construction as long as some aspect
of its form or function is not strictly predictable from its component parts or
from other constructions known to exist. In addition, patterns are stored as
162 Martin Hilpert

constructions even if they are fully predictable as long as they occur with
sufficient frequency.
The starting point of a Collexeme Analysis is the exhaustive collection of
all tokens of a grammatical construction from a corpus. The present study
uses the British National Corpus (Leech 1992) to investigate gerund com-
plement clauses with a to-infinitive, as illustrated in example (9a).
(9) a. Trying to find a job is quite challenging.
With an exhaustive concordance of the gerund clause construction, it can
be determined how frequently a string such as trying to find occurs in it. A
ranked list of the most frequent gerund-infinitive combinations may already
yield a rough impression of the constructional semantics, but a second step
is necessary. The frequency of the string trying to find in the gerund clause
construction has to be compared against its overall frequency in the used
corpus. The basic question that needs to be answered is whether trying to
find is more typical of examples like (9a), or whether it occurs in fact more
often in strings such as (9b) or (9c).
(9) b. I am trying to find a job, and it’s quite challenging.
c. The problem of trying to find a job can be quite challenging.
The typical constructional environment of trying to find can be determined
through a comparison of several frequencies. It first has to be determined
how often trying to find occurs as a gerund clause, and how often it forms
part of a different construction. In the BNC, there are eight examples of
trying to find as a complement clause, and 401 examples in which it occurs
in other constructions. From these numbers, it appears that trying to find is
not a combination that is particularly typical of gerund complement clauses.
Yet, a mere comparison of these two numbers is not enough. It is also im-
portant to know how often other combinations of gerunds and infinitives
occur as complement clauses, and how often they occur in other construc-
tions. The remaining piece of information that is necessary to answer these
questions is the overall frequency of the general pattern V-ing to V. The
BNC contains 52,384 instances of V-ing to V, only 849 of which instantiate
complement clauses. So, even though there are only eight examples of try-
ing to find as complement clauses, these eight examples might be more than
could be expected by chance. In fact, a statistical test such as the Fisher
Exact test shows that trying to find occurs in the gerund complement clause
construction more often than expected, but not significantly so (p=.12).
Table 1 exemplifies the cross-tabulation for such a test.
 Force dynamics of English complement clauses 163

Table 1. trying to find in complement clauses and other constructions

trying to find other combinations totals

complement clauses 8 841 849
other constructions 401 51,134 51,535
totals 409 51,975 52,384

The table contains the information that there are 52,384 instances of V-ing
to V in the BNC, that 849 of these are complement clauses, and that 8 of
these have the form trying to find. It further specifies that the string trying
to find occurs 409 times in the BNC. These pieces of information account
for the figures in the four outer corners of Table 1. The remaining five fig-
ures can be arrived at by subtraction. To calculate association strength be-
tween trying to find and the gerund complement clause construction, the
shaded inner fields of Table 1 are submitted to the Fisher Exact test.
This procedure is not only done for trying to find, but it is repeated for
all combinations of gerunds and infinitives that occur in the retrieved con-
cordance of complement clauses. This was done using the coll.analysis 3.0
script for the statistical processor R (Gries 2004). The result of these calcu-
lations is a list of combinations that are ranked according to their attraction
to the complement clause construction. Attraction here is a function of an
observed frequency that is significantly higher than chance frequency and
therefore results in a small p-value. The smaller the returned p-value, the
more typical a given combination is for the complement clause construc-
tion. Stefanowitsch and Gries (2003) develop the methodology of Collex-
eme Analysis in order to measure the mutual attraction between a gram-
matical construction and individual lexical items. Used in this way,
collexeme allows the study of what kinds of verbs preferably occur in the
English ditransitive construction, or what types of nouns preferably occur
with the English of-genitive. These preferred elements are called collex-
emes, short for collocating lexemes. Since the present study adapts Collex-
eme Analysis to determine preferred strings of words, these strings will be
called collograms, short for collocating n-grams.
The production of a ranked list of collograms is not an end in itself, as
this list still needs to be interpreted. The methodology affords a perspective
on the data that brings the idiosyncrasies of the investigated construction
more into focus than raw frequencies would, but as in an analysis of raw
frequencies, the real work of a semantic study is done by a human analyst.
164 Martin Hilpert

5. A Collexeme Analysis of gerund complement clauses

As discussed in the previous section, the starting point of a Collexeme An-

alysis is the exhaustive extraction of the investigated construction from a
corpus. To this end, all strings of the form V-ing to V are collected from the
BNC, which yields a concordance of 52,384 sentences. The concordance
lines are searched for instances of the target construction, that is, examples
in which the string V-ing to V functions as either subject or object of the
matrix clause. Alphabetical sorting of the left context makes it possible to
categorize the sizeable number of examples with relative ease. Copula
forms such as am, are, and was to the left indicate that the following ge-
rund is not the beginning of a complement clause, as do the prepositions of
and to.
Table 2. The most frequent elements in the English gerund complement clause
construction

Gerund Tokens Infinitive Tokens

having to 210 make 27
being able to 179 read 26
trying to 173 get 20
learning to 92 use 18
attempting to 39 go 14
refusing to 19 see 14
failing to 17 murder 13
getting to 12 find 12
conspiring to 7 put 11
wanting to 6 take 11

Punctuation signs to the left tentatively indicate that the example could be a
complement clause. The identification procedure yields that only 849 ex-
amples of the concordance instantiate the target construction. These exam-
ples contain 67 different gerund types and 358 infinitive types, of which the
ten most frequent are presented in Table 2.
The occurring gerund and infinitive types combine into a total of 499
different types of gerund-infinitive combinations. Table 3 presents the 20
most frequent combination types that are found in the concordance.
 Force dynamics of English complement clauses 165

Table 3. The most frequent combinations in the English gerund complement clause
construction

Combination Tokens Combination Tokens

learning to read 17 having to give 7
having to go 12 having to take 7
trying to make 12 learning to relax 6
being able to see 10 learning to use 6
attempting to murder 9 being able to get 5
getting to know 9 being able to say 5
having to make 9 having to work 5
trying to find 8 trying to kill 5
trying to get 8 trying to read 5
being able to put 7 being able to afford 4

A first analysis based on these raw frequencies allows the following obser-
vations. The most frequent gerunds in Table 2 refer to obligation (having
to) and ability (being able to), thus instantiating deontic modal meanings.
However, it is not possible to reduce the construction to this type of mean-
ing. The gerunds trying and attempting convey the meaning of intention,
failing and refusing as in failing to comply or refusing to eat encode non-
action, and wanting encodes volition. A broader frame of reference is there-
fore needed to capture the constructional semantics, and it will be argued
below that the concept of force dynamics provides such a frame.
With respect to the infinitive complements, the raw frequencies yield a
less coherent picture. The ten most frequent verbs, with the exception of go,
allow for a transitive argument structure. The verb make, read, get, and use
can take a direct object in the investigated construction, as illustrated in the
examples below.
(10) a. Trying to make a will alone can be full of pitfalls. (a-b: BNC)
b. William gave up trying to read the book reviews.
Among the most frequent gerund-infinitive combinations in the investi-
gated construction are highly idiomatic combinations that consist of two
verbs, but arguably express only one concept. The combinations having to
go, getting to know, trying to find , trying to get, learning to relax, learning
to use, being able to say, and being able to afford fall into this category. In
these combinations, the gerund and the infinitive denote a single situation.
The events denoted by the gerund and the infinitive coincide spatio-
temporally. Examples (11a) to (11c) illustrate that the two verbal elements
166 Martin Hilpert

form a conceptual unity that is much stronger than for example the combi-
nation arranging to receive in (11d). Here, receiving goods and the ar-
rangements in order to receive them are two separate events.
(11) a. Getting to know the audience was important. (a-d: BNC)
b. They didn’t like having to go to school.
c. Trying to find somewhere to live in a strange country may also be a
concern.
d. Arranging to receive goods before they have been stolen does not
amount to handling.
Despite the common trait of conceptual unity, the frequent combinations in
Table 3 do not fall into a well-defined semantic frame, as the denoted ac-
tivities cover a wide spectrum. The fact that many combinations refer to a
conceptually integrated event is revealing, and it is consonant with the idea
that nominalized structures portray complex events or entities as unified
wholes (Langacker 1991: 26).
Table 4. Collograms of the English complement clause construction

COLLOGRAM COLLSTR COLLOGRAM COLLSTR

learning to read 22.64 having to take 3.59
being able to see 8.75 having to retype 3.58
being able to put 8.54 being able to experience 3.58
learning to relax 8.11 being able to test 3.58
attempting to murder 6.99 learning to understand 3.33
intending to cause 6.47 being able to pick 3.33
having to go 6.00 being able to write 3.33
learning to use 5.97 being able to cope 3.20
having to give 5.89 having to tell 3.11
being able to say 4.97 conspiring to commit 3.11
having to make 4.95 conspiring to murder 3.11
having to talk 4.88 learning to draw 3.11
learning to see 4.38 learning to sail 3.11
trying to persuade 4.38 having to place 3.11
being able to afford 3.86 having to sacrifice 3.11
having to depend 3.85 being able to employ 3.11
getting to know 3.73 being able to tell 3.11
learning to control 3.65 having to use 3.02

An interesting question to consider is whether this conceptual unity reflects

the constructional semantics of gerund complement clauses or the general
 Force dynamics of English complement clauses 167

association of the combining elements. This question cannot be categori-

cally answered, but it requires a closer look at the individual combinations
and their distribution over different constructions. To this end, all retrieved
combinations from the BNC are submitted to a Collexeme Analysis. Table
4 presents the thirty-six collograms that are significantly attracted to the
construction (p<.001). The table lists eleven gerund types, namely learning,
being able, attempting, having, intending, getting, trying, conspiring, fail-
ing, refusing, and plotting. All of these instantiate the semantic category of
force dynamics, that is they encode the potential or actual exertion of force,
the resistance to it, the clash of two forces, or the absence of a force (cf.
Talmy 2000:409).
The observed gerund types can be grouped into four categories, each of
which involves a particular force-dynamic scenario. The differences be-
tween these scenarios are the result of different inherent force tendencies in
the agonist, and differences with respect to the result of the exerted force.
The following paragraphs analyze each of these four scenarios in detail.

5.1. Attempts

The first force-dynamic category is that of attempts, thus instantiating ac-

tion-referrent gerund clauses (Lees 1960). This category accounts for five
of the eleven gerunds. The gerunds attempting, intending, trying, conspir-
ing, and plotting denote that the agonist exerts force towards the antagonist,
but they are vague with respect to the result of the exerted force. Talmy
(2000: 436) points out that the expression of force-dynamic patterns in any
given language interact with aspectual distinctions, that is, with the presen-
tation of the temporal flow of a situation. The aspectual character of at-
tempts is that they focus on the beginning phase of a force-dynamic scen-
ario, and that they de-focus the end phase of the same scenario. This
differentiates a gerund like attempting from semantically related gerunds
like succeeding or failing, which also involve an attempt, but focus on the
end point of a force-dynamic scenario.
In the below examples from the BNC, the outcome of the agonist’s ac-
tivity is de-focused. It can be pragmatically inferred from example (12a)
that an attempted murder was not completed, but (12b) is completely vague
about the actions of the agonists and their results.
(12) a. Boyce admitted murdering Thomas Boedecker and attempting to mur-
der his wife.
168 Martin Hilpert

b. Both men would deny intending to cause grievous bodily harm.

c. They’d just given up trying to persuade people to dress in 1950s fash-
ions.
The mentioned gerunds co-occur with the infinitives murder, cause, per-
suade, and commit, which are highly transitive (Hopper and Thompson
1980), in that they denote punctual events in which a human agent strongly
affects a patient. The presence of the verbs murder and commit is evidence
that the semantic frame of illegal activity, more specifically the attempt at a
crime, is associated with the gerund complement clause construction. This
association finds a natural motivation in the force-dynamic character of the
construction. Crimes are conceptualized as events that involve the wrongful
exertion of force. In this conceptual schema, criminals are mapped onto the
role of the agonist, who exerts force towards a victim, the antagonist.

5.2. Abilities

The second force-dynamic category concerns abilities. An ability can exist

without ever being applied, and so this type falls into the category of fac-
tive gerund clauses (Lees 1960). Having an ability is a fact, rather than an
action. This category accounts for the eleven combinations in Table 4 that
involve the complex gerund being able to. The defining trait of an ability is
that the agonist has an inherent force that is sufficient to accomplish a re-
sult. However, the agonist’s force is not in effect at all times. Rather, it is
contingent on the agonist’s intention to use it. Talmy (2000: 445) models
the force-dynamic scenario of an ability as a potential force in the agonist,
which is not opposed by an antagonist. Ability is thus re-defined as a poten-
tial force in the absence of blockage. This shows the conceptual closeness
of ability, permission, and prohibition, which all share the notions of a po-
tential force and the presence or absence of blockage to that force. The
examples below illustrate the types of abilities that are expressed with the
investigated gerund complement clause construction.
(13) a. Being able to see both points of view has helped me. (a-e: BNC)
b. Communicating means being able to put a message across and to re-
ceive one.
c. Another example of giving in is not being able to say ‘no’.
d. Being able to afford a dog now is another advantage.
e. It’s being able to pick up the trophy - that’s the nice thing.
 Force dynamics of English complement clauses 169

It was observed earlier that the force-dynamic scenario of attempts was

associated with the semantic frame of illegal activity. The scenario of abili-
ties is associated with very different semantic frames. The infinitives that
co-occur with being able to in Table 4 are see, put, say, afford, experience,
test, pick, write, cope, employ, and tell. Examples (13a) to (13c), as well as
the presence of the verbs write and tell, suggest that the construction is tied
to the semantic frame of communication, as these verbs are used to express
the skills of negotiation and self-expression.
Another group of infinitives, the verbs afford, experience, and pick, en-
code abilities that show the construal of abilities as a commodity. Examples
(13d) and (13e) underscore that the force-dynamic scenario of abilities is
connected to the frame of privilege. In this frame, the agonist is presented
as being able to access a limited and valuable commodity.

5.3. Accomplishments

In the third force-dynamic category, the agonist’s exertion of force leads to

an actual result. This category therefore falls squarely into the action-
referrent type of gerund clauses identified by Lees (1960). Two gerunds
represent the scenario of incremental accomplishments. The gerund getting,
which is only found in the collogram getting to know, entails that an objec-
tive has been accomplished to some degree. Similarly, the gerund learning
describes the incremental progress towards mastery of a skill. As the pro-
gressive form indicates, the construction does not encode that the process is
completed, but that it is still ongoing. The category label of accomplish-
ments, which in itself suggests an end point, is used here because these
activities are viewed as sequences of successful steps, in each of which the
agonist uses force to overcome an antagonistic force and thus accomplishes
an result.
In agreement with this view, examples (14a) to (14c) express the incre-
mental accomplishments that constitute a learning experience. As (14a)
illustrates, even the collogram getting to know falls into the general frame
of learning, which appears to be closely linked to the construction.
(14) a. Getting to know the audience was important. (a-c: BNC)
b. Learning to read involves recognising that writing is made.
c. Learning to use a map and compass is as essential for children as
learning to swim.
170 Martin Hilpert

Both getting and learning co-occur with positively connotated atelic verbs
such as know, read, relax, see, control, draw, and sail. Like the abilities
expressed with being able to, the activities that are expressed in examples
with getting and learning instantiate skills that are conceptualized as valu-
able assets. A semantic characteristic that emerges from the two force-
dynamic scenarios of abilities and accomplishments is that the investigated
gerund clause construction is used to express skills. The only difference
that obtains between the two force-dynamic scenarios is that abilities are
conceptualized as stative facts, whereas the acquisition of an ability is
viewed as a dynamic activity.

5.4. Obligations

The fourth force-dynamic category instantiates obligations, which again fall

into the type of factive gerund clauses (Lees 1960). In the gerund clause
construction, the gerund having instantiates the complex modal auxiliary
have to, which denotes obligations. Obligations represent a more complex
force-dynamic scenario than the previous three, because they involve an
additional type of force. In collograms with having to, a force is exerted on
the agonist with the goal of compelling him to exert force towards an an-
tagonist. The force on the agonist is not physical, but it usually represents
interpersonal social pressure. The pressure is not attributed to a particular
source, but it represents an ambient source of obligation.
Similar to the scenarios of attempts and abilities, it is de-focused
whether any action occurs or any results obtain. Table 4 lists a wide range
of activity verbs that co-occur with having to, such as go, give, make, talk,
take, and tell, but also the stative verb depend. Examples (15a) to (15c)
illustrate that these verbs are used to express common social obligations.
(15) a. They didn’t like having to go to school. (a-c: BNC)
b. Much as she disliked having to talk to Robert, she knew she was
bound to help.
c. Eb resented having to depend on other people to assist him.
In the above examples, individuals experience a negative attitude towards
an obligation that represents a cultural norm. The obligations to comply
with institutions and to act in a socially acceptable way are culturally in-
stalled, such that the actual force behind it cannot be located, but only ex-
perienced.
 Force dynamics of English complement clauses 171

5.5. Inaction

A minor force-dynamic category that is not represented in Table 4, but

which merits discussion because it complements the above-mentioned
categories is the scenario of inaction. This category straddles the distinction
between factive and action-referrent gerund clauses (Lees 1960), because
inaction is a type of action that can be easily construed as a mere fact. In
this category, the agonist’s intrinsic tendency is not to exert force towards
the antagonist. It follows that no results obtain. Gerunds such as failing or
refusing exemplify the category. Both verbs are polysemous, but in con-
junction with an infinitive complement, their meaning is disambiguated
such that they can only refer to inaction. While failing denotes mere inac-
tion, refusing denotes that the agent deliberately chooses not to exert force.
(16) a. He admitted failing to comply with a red signal. (a-b: BNC)
b. Refusing to come to term with reality harms us.
Several gerund-infinitive combinations with the mentioned gerunds denote
events that, like the combinations that were observed with the scenario of
attempts, fall into the semantic frame of illegal activity. In this category
however, the illegal act has been actually accomplished, if only through
inaction.

6. Assessing the predictions of the force-dynamic account

The most important result of the Collexeme Analysis is that the construc-
tional semantics of the investigated gerund clause construction is firmly
rooted in the category of force dynamics (Talmy 2000). The combinations
of gerunds and infinitives that are most typical of the construction have in
common that they involve a constellation of an agonist and an antagonist
that have inherent force tendencies. The five most prominent force-
dynamic scenarios that are expressed by the construction are attempts,
abilities, accomplishments, obligations, and inaction. Each of these scen-
arios has been described in detail.
Beyond the association with these force-dynamic scenarios, the corpus
data suggest that the construction is connected to specific semantic frames.
First, the frame of illegal activity is frequently found in examples that en-
code an attempt or inaction. When abilities and accomplishments are ex-
pressed with the construction, the individual examples often denote skills
172 Martin Hilpert

that are presented as valuable commodities. The construction is also used to

express obligations, in which an ambient force requires the agonist to en-
gage in a pattern of social interaction.
These observations strongly motivate the recognition of gerund clauses
with a to-infinitive complement as a meaningful unit of grammar. The se-
mantic preferences that are reflected in the statistical co-occurrence patterns
suggest that these clauses are in fact a construction in the technical sense
proposed by Goldberg (1995, 2006).
However, a critic might argue that the present study has merely pro-
duced a list of performance phenomena that reveals certain aspects of
usage, but does not yield any insights into speakers’ grammatical know-
ledge (Newmeyer 2004). The following paragraphs are meant to address
this criticism. The claim that speakers associate the investigated construc-
tion with force-dynamic meaning entails several testable predictions. As
will be shown, these predictions are borne out, lending further empirical
evidence to the proposed account.
A first prediction that follows from the force-dynamic account proposed
here is that examples that cannot be interpreted force-dynamically will be
judged as odd, or even ungrammatical. Conversely, an account that does
not make reference to force-dynamic meaning is at a loss to explain why a
collocation that is perfectly grammatical in a regular verb phrase turns out
to be unacceptable in the gerund complement clause construction. For an
illustration, consider (17a) and (17b).
(17) a. The others were all ready and waiting to start again. (BNC)
b. ?Waiting to start again takes a lot of patience.
The collocation waiting to start is fully idiomatic in English. Hence, it
should be possible to use it in a gerund clause. Example (17b) shows that,
while not being outright ungrammatical, the collocation appears unidio-
matic and non-prototypical in that context. A speaker of English can make
sense of the example, but the sentence is unlikely to be found in actual
usage. The present account argues that this is the case because waiting to
start does not involve any force dynamics.
A second prediction of the force-dynamic account concerns polysemous
items. In case of verbal polysemy, the construction should select those
senses of the respective verbs that are force-dynamic. The following exam-
ples show that this is indeed the case.
(18) a. Mr. Spence denies threatening to kill his wife. (BNC)
b. Mrs. Spence was threatening to faint.
 Force dynamics of English complement clauses 173

c. ?Threatening to faint is an unpleasant experience.

The verb threaten can mean ‘announcing the use of force’, as in (18a), but
it can also mean ‘showing alarming signs’, as in (18b). Since only the first
of these meanings is force-dynamic, it is selected in the gerund clause con-
struction. As shown in (18c), the second sense is highly unusual in the ge-
rund clause construction. The idea that being in a state of almost fainting is
an unpleasant experience is not inherently absurd, it is merely the meaning
associated with the syntactic form that leads to the unacceptability of (18c).
An entirely parallel case is illustrated in examples (19a) to (19c). The
verb act has the force-dynamic sense ‘exerting force’, but can also mean
‘perform on stage’. Example (19c) illustrates that the sense of performing is
much less acceptable in the gerund clause construction than the force-
dynamic sense that appears in (19a).
(19) a. Acting to protect your loved ones or property is not murder. (BNC)
b. I tried acting to get famous.
c. ?Acting to get famous didn’t work out.
What the examples in (18) and (19) show is that the gerund complement
clause construction is compatible with the force-dynamic senses of po-
lysemous items, while senses that fall outside the scope of force dynamics
are markedly less acceptable. An account viewing the construction as a
meaningless syntactic template would fail to explain this tendency.
A last prediction of the force-dynamic account has to do with the effect
of semantic coercion. It is one of the key insights of Goldberg (1995) that
constructional contexts can impose markedly different interpretations on
the meaning of sentences than the meaning that is inherent in the compo-
nent lexical items. In sentences such as Mary kissed John unconscious, the
resultative meaning does not stem from a separate lexical entry pertaining
to the verb kiss. Instead, it is the argument structure of the construction
itself that causes the hearer to construe the kissing event in a non-
prototypical way. Coercion effects like these apply to lexical and gram-
matical items that are semantically incompatible with their syntactic con-
texts (Michaelis 2005:51). In a similar vein, it can be argued that the gerund
complement clause construction can impose its meaning on verbs that are
not inherently force-dynamic.
The effect of coercion is illustrated below. In example (20), the colloca-
tion going to see has different meanings in the gerund clause construction
and in a regular verb phrase. Example (20a) can either receive a future in-
terpretation, or it can mean that someone is currently on their way. In each
174 Martin Hilpert

case, the activity is not yet completed. In example (20b), the gerund clause
construction imposes the force-dynamic meaning of an accomplishment on
the collocation. Here, going to see the doctor means that the agonist has
actually completed the journey, and is ready to undergo undressing and a
physical examination.
(20) a. I’m going to see the doctor.
b. Going to see the doctor meant taking your clothes off. (BNC)
The same coercion effect is illustrated by the examples in (21). The collo-
cation remembering to put can refer to a mental effort, as in (21a). In the
gerund clause construction in (21b), the collocation denotes an accom-
plished action. Simply having the necessity of an action on one’s mind does
not avoid the problems that arise if the action is not completed. Yet, this is
what example (21b) literally states. The constructional meaning overrides
this interpretation and imposes the force-dynamic interpretation of an ac-
complishment.
(21) a. He had problems remembering to put on socks before he put on his
shoes.
b. Remembering to put such objects in accustomed places avoids this.
(BNC)
In summary, the force-dynamic account of the investigated gerund com-
plement clause construction is not only supported by the quantitative evi-
dence yielded by the Collexeme Analysis. Its potential to explain the un-
idiomaticity of certain examples, its account of the sense selection with
polysemous verbs, and its prediction of observable coercion effects lend
further empirical support, countering the criticism that usage data are irrel-
evant to the description of grammatical knowledge.

7. The force-dynamic account in the context of previous research

Lastly, it needs to be discussed how the findings of the present study can be
contextualized with the previous approaches to gerund clauses that were
outlined in section 2. The distinction between factive and action-referrent
gerund clauses pointed out by Lees (1960: 58), was found to cut across the
investigated construction in a systematic way, depending on the respective
force-dynamic scenario that is expressed. The categories of attempts and
accomplishments clearly instantiate action-referrent gerund clauses, while
abilities and obligations instantiate the factive type. The observed force-
 Force dynamics of English complement clauses 175

dynamic scenario of inactions is not as easily categorized, thus confirming

Langacker’s point (1991: 32) that speaker construal is the ultimate instance
in the interpretation of a given example.
A different point made by Langacker (1991: 31) does not gain support
from the present analysis. Langacker hypothesizes that the nominal charac-
teristics of gerund complement clauses lead to a bias towards action-
referrent meaning. As was shown above, also the factive type is closely
connected to the investigated construction.
Schachter (1976: 215) argues that factive gerund clauses have a greater
potential to encode definite situations than action-referrent gerund clauses.
This claim is supported by the present study. The obligations that are ex-
pressed by the factive instances of the investigated construction often refer
to definite situations, as can be shown with quantitative evidence. The ob-
served examples with having to have a strong tendency to appear as object
clauses. Out of 212 examples with having to, 181 are object clauses follow-
ing a transitive verb such as avoid, hate, mean, face, and regret, all of
which are negatively connotated. In these examples, individuals experience
a negative attitude towards some obligation. This obligation still represents
a type, but it is clear that it applies in the specific cases of these individuals.
This is consonant with Schachter’s claim that factivity and definiteness are
correlated.
Heyvaert (2003: 222) views factive gerund clauses as denoting atempo-
ralized process types. This view is compatible with the results of the pres-
ent study. Typical factive examples like (22a) and (22b) present a generic
situation type that can not only apply to the individual instance that is being
referred to, but that instantiates a general cultural pattern. The situations of
having to attend school or being able to afford something are culturally
determined frames that form part of speakers’ world knowledge as generic
situation types.
(22) a. They didn’t like having to go to school. (a-b: BNC)
b. Being able to afford a dog now is another advantage.

8. Conclusion

The present study offers quantitative evidence to argue that gerund com-
plement clauses with a to-infinitive are primarily used to express force-
dynamic meanings. It thus identifies complement clauses as another gram-
matical domain besides modality and causation that is pervaded by the se-
176 Martin Hilpert

mantic category of force dynamics. It is further shown that this result is not
merely descriptive, but that it makes correct predictions about grammatical
knowledge, that is, speakers’ intuitions whether a given example is an ac-
ceptable sentence of English or not. Finally, it is shown that quantitative
usage-based approaches (such as Collostructional Analysis) can fruitfully
assess hypotheses about conceptual issues. These points illustrate how
quantitative corpus work can not only be applied to the analysis of Cogni-
tive Semantics, but indeed drive the development of cognitive linguistic
theory.

References

Colleman, Timothy
2009 The semantic range of the Dutch double object construction: A col-
lostructional perspective. Constructions and Frames 1(2): 190-221.
Gilquin, Gaëtanelle
2006 The verb slot in causative constructions. Finding the best fit.
In Constructions All Over: Case studies and theoretical implications
(Special ed. Constructions 1) Doris Schönefeld (ed.), SV1-3/2006
(http://www.constructions-online.de).
Goldberg, Adele E.
1995 Constructions. A Construction Grammar approach to argument
structure. University of Chicago Press.
2006 Constructions at Work: The nature of generalization in language.
Oxford: Oxford University Press.
Gries, Stefan Th.
2004 Coll.analysis 3. A program for R for Windows 2.x.
Heyvaert, Liesbet
2003 A Cognitive-Functional Approach to Nominalization in English.
Berlin / New York: Mouton de Gruyter.
Hilpert, Martin
2008 Germanic Future Constructions. A usage-based approach to lan-
guage change. Amsterdam / Philadelphia: John Benjamins.
Langacker, Ronald W.
1991 Foundations of Cognitive Grammar. Vol. 2. Descriptive application.
Stanford: Stanford University Press.
2005 Construction grammars: Cognitive, radical, and less so. In Cognitive
Linguistics. Internal dynamics and interdisciplinary interaction,
 Force dynamics of English complement clauses 177

Francisco José Ruiz de Menoza Ibáñez and Sandra Peña Cervel

(eds.), 101-159. Berlin / New York: Mouton de Gruyter.
Leech, Geoffrey
1992 100 million words of English: the British National Corpus. Lan-
guage Research 28(1) 1-13.
Lees, Robert B.
1960 The Grammar of English Nominalizations. The Hague: Mouton.
Michaelis, Laura A.
2005 Entity and event coercion in a symbolic theory of syntax. In Con-
struction Grammars. Cognitive grounding and theoretical exten-
sions, Jan-Ola Östman and Miriam Fried (eds.), 45-88. Amsterdam /
Philadelphia: John Benjamins.
Newmeyer, Frederick
2003 Grammar is grammar and usage is usage. Language 79(4): 682-707.
Pullum, Geoffrey K.
1991 English nominal gerund phrases as noun phrases with verb-phrase
heads. Linguistics 29: 763-99.
Quirk, Randolph, Sidney Greenbaum, Geoffrey Leech and Jan Svartvik
1985 A Comprehensive Grammar of the English Language. London:
Longman.
Rubin, Edgar
1958 Figure and ground. In Readings in Perception, David Beardslee and
Michael Wertheimer (eds.), 35-101. Princeton: Van Nostrand. First
published 1915.
Schachter, Paul
1976 A non-transformational account of gerundive nominals in English.
Linguistic Inquiry 7(2): 205-41.
Speelman, Dirk, José Tummers and Dirk Geeraerts
2009 Lexical patterning in a Construction Grammar: The effect of lexical
co-occurrence patterns on the inflectional variation in Dutch attribu-
tive adjectives. Constructions and Frames 1(1): 87–118.
Stefanowitsch, Anatol and Stefan Th. Gries
2003 Collostructions: Investigating the interaction between words and
constructions. International Journal of Corpus Linguistics 8(2): 209-
43.
Talmy, Leonard
2000 Toward a Cognitive Semantics. Cambridge: MIT Press.
Wiechmann, Daniel
2008 Sense-contingent lexical preferences and early parsing decisions.
Cognitive Linguistics 19(3): 439-455.
178 Martin Hilpert

Zeschel, Arne
2008 Lexical chunking effects in syntactic processing. Cognitive Linguist-
ics 19(3): 419-438.
Accounting for the role of situation in language use
in a Cognitive Semantic representation of sentence
mood

Kerstin Fischer

Abstract
Many linguistic forms are part of systems of linguistic choices that differ only in
subtle cognitive-functional respects, which often concern their appropriateness in,
and association with, particular situations. While such relationships have tradition-
ally been described in terms of register, in a usage-based model such aspects have
to be represented as part of their Cognitive Semantic description.
However, situation is not objectively given but individually construed, and thus
both inter- and intrapersonal variation can be found in speakers’ linguistics choi-
ces; the key problem then becomes to identify the speakers’ own cognitive models
of the situation and to relate the occurrence of linguistic features to these models in
order to tease out the subtle cognitive-functional aspects that distinguish the usage
of the linguistic forms of a linguistic subsystem.
In this paper I show the merits and limitations of quantitative and qualitative
methods in the analysis of the role of situation in language use. I present a detailed
study of grammatical mood in a corpus of human-robot interaction that provides an
(objectively) identical situation for all speakers. In the corpus, mood choice can be
shown to be significantly related to the speakers’ different concepts of the human-
robot interaction situation, as well as to other linguistic features indicating differing
cognitive representations of the artificial communication partner. I then propose a
Cognitive Semantic analysis of grammatical mood, using Embodied Construction
Grammar (ECG). This formalism has been designed specifically to account for the
relationship between schematic, extralinguistic knowledge and grammatical choice
(Chang et al. 2002).
Keywords: situation-specific language use, Embodied Construction Grammar,
situation construal, sentence mood, human-robot communication

1. Introduction: Language and situation

In spite of the strong influence the context of situation may have on lan-
1
guage use, situation has so far been hardly dealt with in Cognitive Lin-
180 Kerstin Fischer

guistics – with a few exceptions. For instance, with respect to lexical se-
mantics and idiomatic expressions, Fillmore (1982, 1988, Fillmore et al.
1988) has argued that the semantics of understanding, the semantics Cogni-
tive Linguistics is concerned with, cannot be separated from encyclopedic
knowledge and that semantic representations of lexical items have to in-
clude references to schematic situations. In Frame Semantics (e.g. Fillmore,
Johnson and Petruck 2003), whole scenes with their typical participants are
taken to constitute the meanings of verbs. Similarly, nouns like orphan,
breakfast, weekend or vegetarian (Fillmore 1982) or the days of the week
(Fillmore and Atkins 1992) may have to be viewed in the broad context of
encyclopedic knowledge that provides the frame in which the meanings of
the individual items are located. Thus, the meanings of these items consists
in schematic representations of complex situations.
Recently, Chang and Mok (2006) and Mok and Byrant (in press) have
also proposed models of situation in child-directed speech for pronoun
resolution and argument omission respectively. The authors demonstrate
how situation may be modeled using Embodied Construction Grammar,
describing situational properties as schemata with slots and fillers in which
situation can function as a resource for semantic specification and infer-
ence. Thus, situation functions here as a background resource for the inter-
pretation of utterances.
Another aspect of the relationship between language and situation, that
speakers choose the linguistic features of their utterances on the basis of
what they consider to be situationally appropriate, has however been
largely neglected in Cognitive Semantics. Although cognitive linguists
emphasize the usage-based perspective, this does not generally include
“that kind of usage” (Newman, this vol., emphasis original). This may
have several reasons (cf. Langacker 1999); first of all, cognitive linguists
have concentrated for a long time on the influence of cognition on the
structure, not on the use, of language. Moreover, ubiquitous findings on
categorization show that categories are not objective, but that humans con-
strue them on the basis of cognitive predispositions, embodiment, task and
scene perception, or centrality (e.g. Lakoff 1987, Lakoff and Turner 1989,
Rosch et al. 1976). Such a perspective precludes the simple association of
distributional regularities with given situations. This is in contrast, for in-
stance, to register-based approaches to the relationship between situation
and language use (e.g. Biber 1993, 2006), in which predefined situational
categories are employed and in which the use of quantitative methods is
commonplace: the task is to analyze the probability of a given linguistic
 Accounting for the role of situation in language use 181

feature in a particular situation by means of statistical procedures. Addi-

tional qualitative methods may subsequently be used to identify the func-
tional properties of the linguistic feature that explain its occurrence in that
situation of use (e.g. Biber 2006). However, there is currently no situation
typology which a register description could be based on yet, and while the
notion of ‘situation’ itself suggests an identifiable entity with clear boun-
daries, there is now a considerable body of evidence that situations may
rather be subjectively or interactively construed (Gumperz 1982, Lakoff
1987, Schegloff 1997, Prevignano and di Luzio 2003). That is, while par-
ticular situations may make certain linguistic features conventionally or
functionally relevant, speakers may also employ these features to define the
situation (Tannen 1979). Thus, language use may also contribute to concep-
tualizing the situation. However, if situation is not a given, there is nothing
that language use could be matched with. Finally, recording probabilities of
occurrence does not have much explanatory value by itself. Thus, cognitive
linguists generally insist on providing cognitive-functional explanations for
observable patternings and do not satisfy themselves with recording proba-
bilistic distributional regularities, be they situational, sociolinguistic or
other, for their lack of explanatory value (e.g. Kay and McDaniel 1979,
1981).
We can conclude that although situation has been found to influence
language use quantitatively to a great extent, it has so far been neglected in
cognitive linguistic research, due to the methodological issues described
above, which contradict key cognitive linguistic assumptions. Nevertheless,
cognitive linguistic ideas provide a useful framework for the treatment of
situational influence on language use, for instance, the core assumption that
language construal is subjective; that semantic and encyclopedic know-
ledge cannot be reliably distinguished; that linguistic knowledge and lan-
guage acquisition are usage-based and that language acquisition consists in
a step-wise decontextualisation process (cf. Langacker 1999). Finally,
within the cognitive linguistic tradition, a computational model has been
developed that seems well-suited to account for the relationship between
language and situation, namely Embodied Construction Grammar (Chang
et al. 2002, Bryant 2004, Bergen and Chang 2005, Feldman 2006). Em-
bodied Construction Grammar (ECG) provides a formal specification of the
interaction between central cognitive linguistic concepts, in particular con-
structions, schemata, maps, and spaces.
Constructions are form-meaning pairs that together constitute the
grammar of a language (cf. Goldberg 1995, 2006; Kay 1995, 1997: 123;
182 Kerstin Fischer

Kay and Fillmore 1999; Fillmore 1988). Schemata are representations of all
kinds of schematic knowledge, such as frames or scripts. By maps, ECG
refers to metaphoric and metonymic relationships between domains. And
spaces allow the modeling of blending operations (Fauconnier and Turner
2002). At the same time, ECG assumes that understanding is simulation-
based (Bergen and Chang 2005) and that conceptual and linguistic know-
ledge interact in producing the semantic specification of an expression.
This specification is simulated in a particular situation, which then creates a
rich representation of a scene and allows for numerous inferences to be
drawn.
In the current paper, I provide a model of grammatical mood that is con-
sistent with cognitive linguistic findings, including the subjective nature of
conceptualizations of the situation, and that accounts for the interaction
between linguistic choice and situational features. For that aim, I extend
current work in ECG to capturing situation as a determining factor for lin-
guistic choice. The general properties of ECG, I suggest, provide a unified
formalism for modeling the aspects of situation as well as allowing the
representation of the properties of the conditioning factors of mood choice
identified in the linguistic analysis.

2. Empirical study

2.1. Sentence mood

Sentence mood constitutes a system of linguistic choices that provides dif-

ferent options to reach the same goal. For instance, if your aim is to get
someone to move somewhere, you may say go straight, as in (1). Alterna-
tively, you could say (2)-(10):
(1) um -- go straight?
(2) please go to - goal bowl - number one?
(3) I want you to go to the first object to your right
(4) you should be going to the north-west
(5) the correct object will be the first
(6) the next object that you will go to (2) is (3) three (1) three metres in front
of you
 Accounting for the role of situation in language use 183

(7) could you go towards the (at=prominent)cup(/a) please.

(8) now you should, turn (at=prominent)left(/a)
(9) that cup there. I want you to take please.
(10) do you see the glass object, -- that is, (2) three metres away from you?
The question arising in the face of this variability is what the choice of
sentence mood, for instance, imperative, declarative, interrogative, depends
on. While the trivial answer is that it depends on what the current speaker
considers to be situationally appropriate, the question arises what the situa-
tion consists of and what makes one mood construction more appropriate
than another. These two aspects of the question have usually been dealt
with in isolation; in studies of sentence mood, situation has usually been
treated as a given. The most common proposal is to associate a given mood
construction with a particular speech act. For instance, one might argue that
an imperative clause means something like “I order you”; thus, many
scholars have proposed an intimate relationship between sentence mood
and speech acts (e.g. Halliday 1985, Wierzbicka 1988, Han 2000).
However, there is no direct relationship between speech acts and sen-
tence mood; above we have seen that speakers may use very different mood
constructions to fulfil the same task, to instruct a robot to move somewhere.
Moreover, individual sentence mood constructions can be associated with
different speech acts; for example, as numerous scholars have pointed out
(e.g. Wilson and Sperber 1988, Sbisá 1987), the imperative can be used
with several different speech act functions (cf. Brown and Levinson 1987):
get out (command)
get well (wish)
watch out (warning)
have some tea (invitation)
If however sentence mood is not directly related to a speech act such that
many different mood constructions can be used to carry out the same re-
quest, it is open which factors condition speakers’ choices in a given situa-
tion. The task is thus to tease out the subtle cognitive-functional aspects
that distinguish the usage of the linguistic forms of a linguistic subsystem
depending on the speaker’s construal of the situation. I focus here on the
communicative task of giving instructions and the bases for speakers’ choi-
ces of one instructional strategy over another.
184 Kerstin Fischer

2.2. Methods and data

The procedure taken in this study is to determine first how the participants
themselves conceptualize the situation. This requirement imposes con-
straints on data selection and elicitation. In order to identify participants’
understanding of the situation, the external situation itself should be objec-
tively identical for all participants. For interactive tasks, this is not trivial,
since the coparticipants generally contribute to the definition of the situa-
tion as much as the participant under consideration. However, without
interactive scenarios certain types of data do not occur; for example, certain
sentence moods, such as the imperative, only occur if there is a copartici-
pant.
The solution proposed here is to use human-robot interaction scenarios.
The behavior of a robot as a communication partner can be, unlike other
humans, manipulated at the experimenter’s will. It can be kept identical for
all participants; the robot may even behave independently from the partici-
pants’ utterances. In this case, heterogeneity in participants’ linguistic be-
havior cannot be due to different situational features but has to be attributed
to speakers’ differing preconceptions of what the robot or the human-robot
interaction situation may consist of.
Quantitative analyses determine the influence of the participants’ con-
ceptualization of the situation on language use. In the current analysis, the
question is whether differences in the speakers’ understanding of the situa-
tion correlate with different probabilities of use of the linguistic phenom-
enon under consideration. Quantitative methods allow us to identify statis-
tical relationships between cognitive representations of the situation and
probabilities of linguistic choice.

Participants
The participants in this study were exchange students from various English
speaking countries at the University of Bremen. The corpus consists of
eleven native speakers of English, seven female and four male. Interactions
took about 30-45 minutes.

Data elicitation
The data used here are human-robot interaction dialogs elicited in the
framework of the Collaborative Research Centre Spatial Cognition at the
University of Bremen. The dialogs were elicited in a Wizard-of-Oz scen-
ario (Fraser and Gilbert 1991), in which participants were asked to train a
 Accounting for the role of situation in language use 185

robotic wheelchair on their personal preferences regarding the use of a flat

furnished for a handicapped person. Participants were told that the robotic
wheelchair, the Bremen autonomous robot Rolland (Lankenau et al. 2003),
would learn the labels for particular locations the respective participant
uses to refer to the locations in question. There were four tasks: Task 1
constituted in familiarizing the robot with useful locations in the flat, such
as ‘the kitchen’ but also ‘in front of the TV’. The second task was to sum-
marize the locations taught to the robot from a static position for the robot.
Task 3 was to familiarize the robot with particular locations in the building,
like the room with the photocopying machine. The final task, task 4, was to
instruct the robot to take the user to one of the locations in the building the
robot had been familiarized with previously. Only the last task is instruc-
tional, so that the linguistic analysis will focus on the speakers’ instructions
in this fourth task.
During the teaching situation, participants had to steer the Rolland like a
normal electric wheelchair. Only at the end of the dialogues the robot was
meant to take the user to the location indicated itself. Yet, since the speech-
driven version of Rolland was not available at the time of the experiments,
Rolland would claim that it had to be recharged first and would ask for
fresh batteries.
During the experiments, participants interacted verbally with the robot.
Its utterances were played according to a fixed schema by a human ‘wiz-
ard’ behind the scene. In a pre-study, typical locations, labels and strategies
employed by users had been identified, on the basis of which the schema of
robot utterances for the elicitation of the data used here was designed for
each possible location. The wizard’s task was to check to which location
the user had moved the wheelchair and to play the presynthesized utter-
ances specified for this location in a prespecified order. Although this pro-
cedure seems quite unnatural, the resulting dialogs are in fact quite smooth,
and it is important to understand that for the participants themselves, the
situations are interactional; the perspective that the robot is ‘not really
interacting’ is therefore an etic view taken from the outside. The fact that
all participants are exposed to the same utterances in the same sequence
allows us to compare the dialogues across participants. A questionnaire
study involving pre- and post-questionnaires revealed furthermore that a)
participants had indeed believed to be talking to an automatic speech-
processing system, b) they found the interactions pleasant, and c) their atti-
tude towards robots in general improved significantly during the interac-
186 Kerstin Fischer

tions (Andonova 2006). The following sample dialog illustrates the relative
smoothness of the interactions:
(11) R022: to go to the stugaroom, (1) turn, (4) onehundred’neighty degrees,
(2), exit, -- room three four four zero. turn, -- left. (3) go straight, (1)
ahead.
Rolland: up to which point do I go straight?
R022: to:: the:: I don’t know, (1) um (4) what’s the place there, (6) go
straight, - turn left to the, (3) to the copy room. go s+
Rolland: When I’m left of the room with the photocopying machine, what
do I do?
R022: you: go: straight until, you get to the stugaroom.
Rolland: okay, I think I got it. I’m taking you to the stugaroom now.
R022: okay, (5) go,
Rolland: I’m sorry, my batteries are too low. Please recharge them first.
Data coding
The subjective construal of the situation needs to be operationalized in
some way. My suggestion is to use the dialog openings since these are
places that indicate speakers’ preconceptions about their artificial com-
munication partner as well as their definition of the situation. Moreover,
previous studies have shown that the users’ attitude towards the system is
the most influential factor in the predictability of user behavior throughout
the dialogs, much more influential than sociolinguistic variables or scenario
differences (Fischer 2006ab).
The dialogs were manually coded for dialog beginnings in the following
way:
The first utterance of the robot is always ‘yes, hello, how do you do?’.
This utterance allows relevant contributions at different levels: speakers can
react to the greeting (e.g. providing ‘hello’), to the question (e.g. by reply-
ing to it with, for instance, ‘fine’), to the polite function of the question
(e.g. by replying ‘thank you’), or to the pragmatic act as a whole (e.g. by
responding with the counter question ‘and how do you do?’). Dialog begin-
nings were now simply coded for the number of strategies employed by
each speaker; the coding thus provides the raw score. For instance, the dia-
log beginning for R004 was coded as 0:
(12) Rolland: Yes, hello, how do you do?
R004: (4)
Rolland: You can take us now to a place you want to name.
R004: (2) table. (laughter)
 Accounting for the role of situation in language use 187

R017 was coded as 1 for the minimal reaction to the content of the question
(I'm good):
(13) Rolland: Yes, hello, how do you do?
R017: (2) I’m good,
Rolland: You can take us now to a place you want to name.
R017: (3) I would like to go::: to the::: computer. straight.
R051 was coded as 2, since the speaker answers the question (fine) and
recognizes the polite function of the robot’s question (thanks):
(14) Rolland: Yes, hello, how do you do?
051: (laughter) fine thanks. okay so,
Rolland: You can take us now to a place you want to name.
R051: (1) we are now going to – the table.
R043 was coded as 3 since the speaker reacts to the contents of the ques-
tion (I'm fine), to its polite function (thank you), and she reciprocates it
(hello, how are you):
(15) Rolland: Yes, hello, how do you do?
R043: (1) hello. (1) I’m fine thank you how are you?
Rolland: You can take us now to a place you want to name.
R043: -- wait a minute, I haven’t finished reading (...) okay, um -
Rolland: You can take us now to a place you want to name.
R043: (2) do you see the plant on the left forty five degrees, a green
plant?
These different dialog beginnings reflect the considerable differences in
how speakers understand the situation in terms of the relationship between
the participants; while some speakers reply to the polite greeting of the
robot, others do not react to the social aspects of the robot’s messages at all.
Speakers therefore conceptualize the situation differently, varying in the
degree to which they understand the situation as social. The most suitable
interpretation of these findings is that speakers may, or may not, enter a
level of joint pretense (Clark 1996, 1999; cf. Fischer 2006a).
Moreover, the instructions from the instructional fourth task of the dia-
logs were coded for the instructional strategy chosen. The variable com-
prises the grammatical moods declarative, imperative, and interrogative. In
addition, speakers also produced instructions without overt verbs; instead
we find adverbial phrases, prepositional phrases and just noun phrases by
means of which speakers instruct the robot. In particular, the following
instructional strategies were distinguished:
188 Kerstin Fischer

Example
declarative R013: so we are going out of , - this room,
imperative R017: turn around and leave the room
adverbial phrase R017: straight,
prepositional phrase R048: to the sofa?
noun phrase R004: table.
interrogative R043: (2) do you see the plant on the left
forty five degrees, (1)
In addition, dialogs elicited in the instructional fourth task were manually
coded for several linguistic properties, covering a spectrum of morphosyn-
tactic, lexical and pragmatic features. Features were chosen based on their
relative frequency; moreover, previous research has shown them to be good
indicators of the communication partner’s suspected competence. Each
users’ linguistic behavior was coded in the following way:
1. number of structuring cues (see also Fischer and Bateman 2006),
comprising implicit (for instance, now or so) and explicit (for in-
stance, first of all, the next step or and then) structuring cues; the in-
dividual tokens were counted for each speaker and, since their occur-
rence is relevant for the relationship between utterances, divided by
the number of utterances;
2. number of relative clauses used divided by the number of utter-
ances; relative clauses have previously been identified as indicators
of high ascribed competence (cf. Fischer 2006a);
3. number of politeness formulas, such as please, thank you or sorry
divided by the number of utterances of each speaker;
4. the number of utterances uttered by each participant in the task
under consideration; the number of utterances tells us about speak-
ers’ linguistic effort spent on the instruction.
The statistical description and analysis was carried out using the Statistica
software package.

2.3. Results of the empirical analysis

The task of the statistical analysis is to show that the different conceptuali-
zations of the situation as indicated by the different dialog openings corre-
 Accounting for the role of situation in language use 189

late with different linguistic behaviors, including different choices of

grammatical mood. Therefore, the correlation between dialog openings and
linguistic features was calculated.
Table 1 shows the average frequency, range and standard deviation (sd)
for the features investigated. The table shows that the imperative is the
most frequent construction in this instructional task for many speakers, as
can be expected from the literature in which the imperative is associated
with requesting. Yet other instructional strategies are also frequent as well.
Just the instruction by means of nouns and adverbs was so infrequent (in
contrast to other corpora of human-robot interaction, see Fischer 2006a)
that even combined they make up only 7% of the utterances per speaker.
Table 1. Descriptive statistics for instructional strategies; n = 11
mean range Sd
Imperative 0.456320 0% - 100% 0.288339
Declarative 0.441694 0% - 80% 0.271049
Prepositional Phrase 0.065754 0% - 25% 0.085747
Interrogative 0.014354 0% - 16% 0.047607
Noun/Adverb 0.011778 0% - 7% 0.026761

Table 2 shows the mean, range and standard deviation for dialogue begin-
nings and for structuring cues, relative clauses, politeness formulas and
number of utterances for the eleven participants.
Table 2. Descriptive statistics for dialog openings and other features; n = 11
mean range Sd
structuring cues 0.642634 0 - 1.33 0.471420
relative clauses 0.041270 0 - 0.2 0.073497
politeness formula 0.062839 0 - 0.33 0.106116
utterances 10.1818 2 - 19 4.729021
dialogue beginning 2.090909 0-3 1.044466

Quantitative analyses of the data using a Pearson product-moment correla-

tion matrix show that the different conceptions of the human-robot situation
as either social or nonsocial correlate significantly with the choice of a
broad spectrum of different linguistic features, including sentence mood.
The analysis reveals a significant negative correlation between the de-
clarative and the imperative mood, indicating that speakers who tend to use
190 Kerstin Fischer

the one do not use the other. Moreover, the declarative is significantly cor-
related with social concepts of the human-robot interaction situation as
indicated by the high attention to social aspects of communication in the
dialog openings. Conversely, the imperative mood is negatively associated
with the social aspects of dialog beginnings.
Table 3. Correlations between the features investigated; * = p < .05, n = 11;
Imp Decl PP Interr N/adv struct relcl polite begin utts
Imp -0.93* -0.35 -0.10 0.10 -0.28 -0.33 -0.31 -0.70* -0,59
Decl -0.93* 0.07 -0.03 -0.32 0.25 0.46 0.21 0.72* 0.35
PP -0.35 0.07 -0.25 0.14 0.12 -0.28 -0.09 -0.01 0.40
Interr -0.10 -0.03 -0.25 0.51 -0.16 -0.19 0.30 0.29 0.62*
N/adv 0.10 -0.32 0.14 0.51 -0.11 -0.27 0.01 -0.13 0.54
struct -0.28 0.25 0.12 -0.16 -0.11 0.50 0.31 0.34 0.09
relcl -0.33 0.46 -0.28 -0.19 -0.27 0.50 0.37 0.21 -0.18
polite -0.31 0.21 -0.09 0.30 0.01 0.31 0.37 0.27 0.19
begin -0.70* 0.72* -0.01 0.29 -0.13 0.34 0.21 0.27 0.64
utts -0.59 0.35 0.40 0.62* 0.54 0.09 -0.18 0.19 0.64*

Legend: Imp = imperative, Decl = declarative, PP = prepositional phrase, Interr = interroga-

tive, N/adv = noun and adverbial instruction, struct = structuring cues, relcl = relative
clauses, polite = politeness formulas, begin = dialog openings, utts = number of utterances in
task 4.

The declarative is furthermore positively associated with the number of

turns and relative clauses, relationships that however do not reach statistical
significance. The imperative mood in contrast is associated with rather
short dialogs and with fewer relative clauses and politeness formulas.
The type of greeting correlates significantly with the declarative (posi-
tively) and the imperative (negatively), as well as with the number of utter-
ances produced, indicating that those who regard the human-robot interac-
tion situation as more social also invest more effort in the dialogs.
Furthermore, there is a tendency for speakers who greet the robot to use
structuring cues, interrogative clauses and politeness formulas.
To sum up, speakers who use the declarative judge the robot’s linguistic
and social capabilities higher than those who use the imperative. The condi-
tions of use for the declarative comprise the understanding of the situation
as social, reciprocal and solidary. In contrast, for the imperative there is no
such correlation. So the quantitative analysis reveals interdependencies
between linguistic features, which in turn provide information on the subtle
cognitive-functional differences between the different mood constructions.
 Accounting for the role of situation in language use 191

To conclude, the analysis presented has shown that although the task
participants are presented with has traditionally been associated with the
imperative mood, several different grammatical mood constructions are
being employed; in their grammatical choices, speakers rely on aspects of
the situation present, such as the task, which is taken to be common ground
between both interactants; the variability in mood choice corresponds to
participants’ differing conceptionalizations of the situation and their com-
munication partner as social or non-social, as reciprocal and solidary or as
non-reciprocal and nonsolidary. These findings have several implications
for a cognitive linguistic model of sentence mood:
1. it needs to rely on conceptual representations of the situation (a situ-
ational frame or schema);
2. it needs to represent the general understanding of the task as com-
mon ground;
3. it needs to account for the interaction between concept of situation
and grammatical choice.

3. Model

The proposal I want to make is to represent a general aspect of the situa-

tion, namely the task, which has been set by the experimenters, as a
schema. That is, the task really seems to be a given for all participants.
None of them attempted to inform the robot what the task consists of and
what the interaction will be about. On the basis of the givenness of this
task, the use of adverbs, PPs and noun phrases as requests to move some-
where can be explained. That is, only if it is situationally available for all
participants that the speaker wants the addressee to move somewhere, an
instruction like “a turn to the left” can be identified as an instruction. Thus,
the first element of the model proposed will be the S wants A to X
Schema.
This schema corresponds to the idealized cognitive model proposed by
Panther and Thornburg (1998) who argue that requests can be understood
as scenarios with sequentially ordered parts, such that there are precondi-
tions, coded in the before-component, the core-component and an after-
component, describing the result. Panther and Thornburg’s point is that the
before- and after-components may metonymically stand for the core-
component, thus giving rise to indirect speech acts, for instance, can you
pass the salt (question about the ability, i.e. precondition of the before-
192 Kerstin Fischer

component) or now you walk straight (description of the result, i.e. an as-
pect of the after-component).
Before: S wants A to do X
A is able to do X
Core: S puts A under obligation to do X
Result: A is under obligation to do X
After: A will do X.
2
This model can be transferred into an ECG representation:
schema SwantsAtoX
subcase of Wanting
evokes SpeechSit as s
roles
before
core
after
constraints
wanter ↔ s.Speaker
wanted-of ↔ s.Addressee
before ↔ ability
core ← “S wants A to do X”
after ← “A will do X”
This schema relates to two more general schemata. On the one hand, the S
wants A to X schema constitutes a subcase of Wanting. On the other, it
evokes the speech situation schema in which the roles of speaker and ad-
dressee are defined, as well as the possible relationship between the two:
schema Wanting
subcase of Action
roles
wanter : Human
wanted-of : Human
wanted : Event
costs
benefit
ability
 Accounting for the role of situation in language use 193

schema SpeechSit
subcase of Schematic-Form
roles
Speaker
Addressee
Relationship
The schemata are taken to be shared by all participants, and thus they are
inherited into participants’ personal situation models (PSMs). These models
however differ with respect to the perception of the interpersonal relation-
ship between human user and robot. Thus, while some speakers pretend the
situation to be like normal conversation (henceforth: the ‘players’), others
do not enter this level of joint pretense (Clark 1999). Since we had been
able to identify user groups for which these assumptions hold, we can take
these situation models to be partly schematic, but ECG would in principle
also allow personal, idiosyncratic models of the situation.
schema PSMPlayer
subcase of SpeechSit
constraints
Relationship ← reciprocal, solidary
In contrast to the ‘players’, there are ‘non-players’ who will define and
understand the human-robot situation as non-reciprocal and as non-
solidary.
schema PSMNonplayer
subcase of SpeechSit
constraints
Relationship ← non-reciprocal, non-solidary
As described above, Panther and Thornburg (1998) hold that the subparts
of the request-scenario may serve instead of others, in particular, that be-
fore- and after-components may give rise to indirect speech acts. In ECG
this can be represented in maps, which specify which roles from which
3
schemata may be combined in source-target pairs, for instance:
map After-for-Core
evokes SwantsAtoX as w
roles
source → w.after
target → w.core
pairs
w.after → w.core
194 Kerstin Fischer

Further components of the model are self-evidently the different sentence

4
mood constructions and their instantiations as requests. The model pro-
posed reflects Langacker’s (2008) position that the grammatical moods
differ with respect to the construals that invoke discourse participants in
different ways. The declarative, for instance, refers to the epistemic level,
that is, it describes conceptual content (2008: 474). Thus, in order for the
declarative to work as a request, it describes the result of a request scenario,
which is modeled here as the mapping between the after-component and the
core-component.
construction DeclRequest
constructional
constitutents
agt : Ref-Expr
v : Verb
form
agt.f before v.f
meaning
evokes SpeechSit as s
evokes SwantsAtoX as w
evokes AfterforCore as a
evokes PSMPlayer as p
v.m.agent ↔ s.Addressee
v.m ↔ w.core
s.Relationship ↔ p.Relationship
In contrast, the imperative construction encodes no reciprocal or solidary
relationship with the communication partner. When this construction inter-
acts with the situational concept, the result is a construction that can be
used for making requests and which evokes the non-player situation model,
avoiding any interpersonal commitments:
construction ImperativeRequest
constructional
constitutents
v : Verb
form
v.f
meaning
evokes SpeechSit as s
evokes AfterforCore as a
 Accounting for the role of situation in language use 195

evokes SwantsAtoX as w
evokes PSMNonplayer as p
v.m.agent ↔ s.Addressee
v.m ↔ w.core
s.Relationship ↔ p.Relationship
Since the two different mood constructions encode links to particular situa-
tion models, use of these constructions can contribute to the definition of
the situation as social, reciprocal and solidary or not. Thus, the model ac-
counts not only for the choice of sentence mood in a given situation but
also explains how the use of a given construction can contribute to the situ-
ational construal as well.

4. Conclusion

To sum up, we have seen how both qualitative and quantitative corpus ana-
lyses can be helpful for the creation of Cognitive Semantic representations
involving situationally determined language use. Even though Cognitive
Linguistics does not assume objectively-given categories, quantitative and
statistical analyses of situation-specific use can be useful if they are appro-
priately combined with qualitative investigations.
Moreover, it was shown how Cognitive Semantic concepts, such as con-
structions, schemata, and metonymic mappings, may interact to account for
language use appropriate for a situation as it is conceived of by the partici-
pants themselves. Thus, we were able to specify interactions between dif-
ferent types of information in accordance with the major principles of Cog-
nitive Linguistics. These interactions provide useful extensions to the ECG
model, which has proven suitable for the flexible modeling of the empirical
findings of this study on situation as a factor influencing language use and
which previously had been used to model situation as a resource, for in-
stance, for pronoun resolution (Chang and Mok 2006, Mok and Bryant in
press).
We can conclude that the influence of situation on language use can be
investigated by means of quantitative methods if there is an orientation
towards speakers’ own understanding of the situation. Thus qualitative and
quantitative methods can be usefully combined. In the case of understand-
ing the subtleties of functions of particular sentence moods, quantitative
methods proved crucial for identifying the interdependencies between par-
196 Kerstin Fischer

ticular linguistic choices that allow inferences about the cognitive models
of the situation evoked by the participants.

Notes

1. For instance, Biber (1993, 2006; Biber et al. 1998) has shown that register
functions as a reliable predictor of language use, as much as language use
predicts register.
2. With the exception of the after-for-core map, which is not yet implemented,
the model presented here was tested for formal correctness using John Bry-
ant’s construction analyzer (Bryant 2004). In addition, since before constitutes
a keyword in the analyzer, the spelling of the before slot in the request-
scenario had to be changed temporarily.
3. The After-for-Core-Map is of course not the only possible mapping; for in-
stance, also the wish itself (as in ‘I’d like you to go straight’) or the ad-
dressee’s ability (as in ‘Can you pass the salt?’) can metonymically stand for
the request (cf. also Ruiz de Mendoza Ibáñez and Baichi 2006).
4. In this paper, the focus is on the interaction between situational features and
the semantics of constructions. For a more comprehensive CxG account of
grammatical mood, see Stefanowitsch (ms.).

References

Andonova, Elena
2006 On changing mental models for a wheelchair robot. In Proceedings
of the Workshop on ‘How People Talk to Computers, Robots, and
Other Artificial Communication Partners’, Hansewissenschaftskol-
leg, Delmenhorst, April 21-23, 2006, Kerstin Fischer (ed.), 131-139.
SFB/TR8 Report 010-09_2006.
Bergen, Benjamin and Nancy Chang
2005 Embodied Construction Grammar in simulation-based language
understanding. In Construction Grammars: Cognitive grounding
and theoretical extensions, Jan-Ola Östman and Miriam Fried (eds.),
147-190. Amsterdam / Philadelphia: John Benjamins.
Biber, Douglas
1993 Using register-diversified corpora for general language studies.
Computational Linguistics 19(2): 219-241.
 Accounting for the role of situation in language use 197

2006 Dimensions of Register Variation: A cross-linguistic comparison.

Cambridge: Cambridge University Press.
Biber, Douglas, Randi Conrad and Susan Reppen
1998 Corpus Linguistics: Investigating language structure and use. Cam-
bridge: Cambridge University Press.
Brown, Penelope and Stephen Levinson,
1987 Politeness. Some universals in language usage, 2nd ed. Cambridge:
Cambridge University Press.
Bryant, John
2004 Scalable construction-based parsing and semantic analysis. In Pro-
ceedings of Second Workshop on Scalable Natural Language
Understanding, Robert Porzel (ed.), 33-40. Boston, Mass.: Associa-
tion for Computational Linguistics.
Chang, Nancy, Jerome Feldman, Robert Porzel and Keith Sander
2002 Scaling Cognitive Linguistics: Formalisms for language understand-
ing. Proceedings of the First Workshop on Scalable Natural Lan-
guage Understanding. <http://citeseerx.ist.psu.edu/viewdoc/
download?doi=10.1.1.9.1170&rep=rep1&type=pdf>
Chang, Nancy and Eva Mok
2006 A structured context model for grammar learning. International
Joint Conference on Neural Networks, 16-21 July 2006, Vancouver,
BC, Canada.
Clark, Herbert H.
1996 Using Language. Cambridge: Cambridge University Press.
1999 How do real people communicate with virtual communication part-
ners? In Proceedings of the Sixteenth National Conference on Artifi-
cial Intelligence, November 5-7th, 1999, North Falmouth, MA.
Ramasamy Uthurusamy and Barbara Hayes-Roth (eds.), 43-47.
Menlo Park, Calif.: AAAI Press.
Fauconnier, Gilles and Mark Turner
2002 The Way We Think: Conceptual blending and the mind’s hidden
complexities. New York: Basic Books.
Feldman, Jerome A.
2006 From Molecule to Metaphor. A neural theory of language. Cam-
bridge, Mass / London: MIT Press.
Fillmore, Charles J.
1982 Frame semantics. In Linguistics in the Morning Calm, Linguistic
Society of Korea (ed.), 111-138. Seoul: Hanshin.
1988 The mechanisms of ‘Construction Grammar’. In Proceedings of the
Fourteenth Annual Meeting of the Berkeley Linguistic Society, S.
Axmarker, A. Jassier, and H. Singmaster (eds.), 35-33. Berkeley,
CA: BLS
198 Kerstin Fischer

Fillmore, Charles J., Paul Kay and Mary O’Connor

1988 Regularity and idiomaticity in grammatical constructions: The case
of let alone. Language 64(3): 501–538.
Fillmore, Charles and Beryl T. Atkins
1992 Towards a frame-based lexicon: The case of RISK. In Frames,
Fields, and Contrasts: New essays in semantic and lexical organiza-
tion, Adrienne Lehrer and Eva F. Kittay (eds.), 75-102. Mahwah,
N.J.: Erlbaum.
Fillmore, Charles J., Christopher R. Johnson and Miriam R.L. Petruck
2003 Background to Framenet. International Journal of Lexicography
16(3): 235-250.
Fischer, Kerstin
2006a What Computer Talk Is and Is not: Human-computer conversation
as intercultural communication. (Linguistics - Computational Lin-
guistics 17.) Saarbrücken: AQ.
2006b The role of users’ preconceptions in talking to computers and robots.
In Proceedings of the Workshop on ‘How People Talk to Computers,
Robots, and Other Artificial Communication Partners’ Hansewis-
senschaftskolleg, Delmenhorst, April 21-23, 2006, Kerstin Fischer
(ed.), 112-130. SFB/TR8 Report 010-09_2006.
Fischer, Kerstin and John A. Bateman
2006 Keeping the initiative: An empirically motivated approach to pre-
dicting user-initiated dialogue contributions in HCI. In Proceedings
of EACL ‘06, 185-192.
Fraser, Norman M. and Nigel G. Gilbert
1991 Effects of system voice quality on user utterances in speech dialogue
systems. In Proceedings of Eurospeech 1991, 57-60.
Goldberg, Adele
1995 Constructions: A Construction Grammar approach to argument
structure. Chicago: University of Chicago Press.
2006 Constructions at Work. The nature of generalization in language.
Oxford: Oxford University Press.
Gumperz, John
1982 Discourse Strategies. Cambridge: Cambridge University Press.
Halliday, Michael A.K.
1985 Introduction to Systemic Functional Grammar. London: Arnold.
Han, Chung-Hye
2000 The Structure and Interpretation of Imperatives: Mood and force in
universal grammar. (Outstanding Dissertations in Linguistics
Seires.) New York: Garland.
 Accounting for the role of situation in language use 199

Kay, Paul
1995 Construction grammar. In Handbook of Pragmatics, Jef Ver-
schueren, Jan-Ola Östman, Jan Blommaert and Chris Bulcaen (eds.),
171-177. Amsterdam / Philadelphia: John Benjamins.
1997 Construction grammar. In Words and the Grammar of Context, Paul
Kay (ed.), 123-131. Stanford: CSLI.
Kay, Paul and Chad K. McDaniel
1979 On the logic of variable rules. Language in Society 8: 151-87.
1981 On the meaning of variable rules. Language in Society 10: 251-58.
Kay, Paul and Charles J. Fillmore
1999 Grammatical constructions and linguistic generalizations: The
What’s X doing Y? construction. Language 75(1): 1-33.
Lakoff, George
1987 Women, Fire, and Dangerous Things. What categories reveal about
the mind. Chicago / London: The University of Chicago Press.
Lakoff, George and Mark Turner
1989 More than Cool Reason: A field guide to poetic metaphor. Chicago /
London: University of Chicago Press.
Langacker, Ronald W.
1999 The contextual basis of Cognitive Semantics. In Language and Con-
ceptualization, J. Nuyts and E. Pederson (eds.), 229-259. Cam-
bridge: Cambridge University Press.
2008 Cognitive Grammar: A basic introduction. New York: Oxford Uni-
versity Press.
Lankenau, Axel, Thomas Röfer and Bernd Krieg-Brückner
2003 Self-localization in large-scale environments for the Bremen au-
tonomous wheelchair. In Spatial Cognition III (Lecture Notes in Ar-
tificial Intelligence 2685), Christian Freksa, W. Brauer, Christopher
Habel and K. F. Wender (eds.), 34-61. Berlin: Springer.
Mok, Eva and John Bryant
In press A best-fit approach to productive omission of arguments. (to ap-
pear). Proceedings of the 32nd Annual Meeting of the Berkeley Lin-
guistics Society: General session and parasession on theoretical ap-
proaches to argument structure. Zhenya Antić, Charles B. Chang,
Jisup Hong, Michael Houser, Maziar Toosarvandani and Yao Yao
(eds.). Berkeley, CA: BLS.
Newman, John
this vol. Balancing acts: Empirical pursuits in Cognitive Linguistics.
Panther, Klaus-Uwe and Linda Thornburg
1998 A cognitive approach to inferencing in conversation. Journal of
Pragmatics 30: 755-769.
200 Kerstin Fischer

Prevignano, Carlo L. and Aldo di Luzio

2003 A discussion with John J. Gumperz. In Language and Interaction:
Discussions with John J. Gumperz, Susan L. Eerdmans, Carlo L.
Prevignano and Paul J. Thibault (eds.), 7-29. Amsterdam / Philadel-
phia: John Benjamins.
Rosch, Eleanor H., Carolyn B. Mervis, Wayne D. Gray, David M. Johnson and
Penny Boyes-Braem
1976 Basic objects in natural categories. Cognitive Psychology 8: 382-
439.
Ruiz de Mendoza Ibáñez, Francisco José and Anna Baichi
2006 Illocutionary Constructions: Cognitive motivation and linguistic
realization. Essen: L.A.U.D. A 668.
Sbisá, Marina
1987 Indicative Mood, Illocutionary Force, and Truth: Some points for
discussion. Duisburg: LAUD. A 193.
Schegloff, Emanuel A.
1997 Whose text, whose context? Discourse & Society 8(2): 165-187.
Stefanowitsch, Anatol
ms The English imperative: A construction-based approach. Unpub-
lished manuscript.
Tannen, Deborah
1979 What’s in a frame? Surface evidence for underlying expectations. In
New Directions in Discourse Processing, R. Freedle (ed.), 137-181.
Norwood, N.J.: Ablex.
Wierzbicka, Anna
1988 The Semantics of Grammar. Amsterdam / Philadelphia: John Ben-
jamins.
Wilson, Dan and Deirdre Sperber
1988 Mood and the analysis of non-declarative sentences. In Human
Agency: Language, duty and value, Jonathan Dancy, J. Moravcsik
and C. Taylor (eds.), 77-101. Stanford, CA.: Stanford University
Press.
Exemplars and analogy:
Semantic extension in constructional networks

Arne Zeschel

Abstract
How (and when) do speakers generalise from memorised exemplars of a construc-
tion to a productive schema? The present paper presents a novel take on this issue
by offering a corpus-based approach to semantic extension processes. Focusing on
clusters of German ADJ N expressions involving the heavily polysemous adjective
tief ‘deep’, it is shown that type frequency (a commonly used measure of produc-
tivity) needs to be relativised to distinct semantic classes within the overall usage
spectrum of a given construction in order to predict the occurrence of novel types
within a particular region of this spectrum. Some methodological and theoretical
implications for usage-based linguistic model building are considered.
Keywords: semantic extension, productivity, type frequency, Correlation Analysis,
usage-based model

1. Introduction

How (and when) do speakers generalise from specific learned expressions

1
of a particular type to a productive constructional schema? Usage-based
approaches to language generally emphasise that these two notions are in
fact endpoints on a continuous scale. Still, positing a blurred boundary be-
tween lexical and grammatical constructions does not in itself explain how
speakers derive the latter from the former. For instance, consider the status
of the putative schema in (1.a) in view of the acceptability of the expres-
2
sions in (1.b-d):
(1) a. to go ADJ
b. to go mad / bankrupt / quiet …
c. to go crazy / insane / mental … broke / bust / ?skint… calm / silent / ?still…
d.to go ?famous / ?popular / ??renowned …
In a paper about a group of similar constructions in Spanish, Bybee and
Eddington (2006) argue that acceptability reflects familiarity: an expression
is acceptable to the extent that it is a.) directly retrievable from memory as
202 Arne Zeschel

a stored unit (as an effect of high input frequency), or b.) semantically simi-
lar to such a stored unit.
This account offers a both simple and intuitively plausible explanation
for the differences in idiomatic feel or ‘naturalness’ between the routine
formulas in (1.b), their close semantic variants in (1.c), and the less obvi-
ously related verb-adjective combinations in (1.d). As it stands, however, it
has little to say about the status of (1.a): on the one hand, the dubious ac-
ceptability of (1.d) suggests that the adjective position in this pattern is not
a fully general ‘open slot’. On the other hand, the verb go nevertheless does
occur with several dozens of adjectives in such change-of-state expressions
in the British National Corpus, and many of these combinations have a
distinctly unconventional ring to them (suggesting that they are in fact cre-
ative formations that were computed from scratch rather than stored and
retrieved ‘fixed expressions’). So where do these come from? Since a fully
general productive schema [go ADJ] does not seem to exist, it would ap-
pear that such combinations are formed in analogy to specific existing ex-
pressions of the relevant type, and this is indeed what Bybee and Eddington
conclude from an analysis of their Spanish data. However, this answer
raises an important new question: if semantic extension proceeds via an-
alogy to entrenched exemplars, then how is it that some high-frequency
expressions are easily extended in this way (cf. 2.a), whereas others permit
considerably less or even little to no such variation at all (cf. 2.b, c)?
(2) a. to go mad / crazy / insane / wild / furious / ballistic …
b. to go bankrupt / broke / bust / ?skint / ?insolvent / ?illiquid …
c. to go cold / ?cool / ?freezing / ?icy / ?chilly / ?frosty …
One idea is that the crucial additional property is semantic type frequency
(Israel 1996): on this account, the likelihood (in production) and accept-
ability (in comprehension) of a particular novel extension depends on the
amount of similar expressions that are already known to the speaker. The
present paper evaluates this hypothesis with a quantitative corpus study of a
group of idiomatically restricted collocations. Focusing on figurative adjec-
tive-noun expressions involving the heavily polysemous adjective tief
(‘deep’) in German, it introduces a corpus-based approach to the recon-
struction of semantic extension processes in constructional networks. The
prediction to be evaluated is the following:
(3) Semantic type frequency hypothesis
The likelihood/acceptability of a particular novel extension of an established
collocation depends on the number of previously encountered variants of the
same semantic type
 Exemplars and analogy 203

The paper is structured as follows: I begin with a brief sketch of the

theoretical motivations for the above assumption (section 2) and introduce
methods and data of the study (section 3). Section 4 evaluates the hypoth-
esis on three successively more detailed levels of semantic analysis. Section
5 offers a brief summary and conclusion.

2. Semantic extension: A usage-based account

Especially in morphology, type frequency has long been recognised as a

useful measure of productivity (Bybee 1985, 1995): on this account, a good
indication that e.g. affixation with –s is a more productive strategy for ex-
pressing the meaning PLURAL in English than affixation with –en is that
there are substantially more plural formations (types) that involve an allo-
morph of –s (e.g. box > boxes) than in the case of –en (e.g. ox > oxen).
Since usage-based approaches see no categorical difference between mor-
phology and syntax in principle (in the sense that they involve qualitatively
different representations or processing operations), an extension of these
findings to word combinations in syntax would appear natural. On the other
hand, even in a model that sees lexicon and syntax as a continuum, what is
a useful measure for a highly generalised schema such as [N-s] may well be
less appropriate for a much less inclusive (and in fact maybe just hypotheti-
cal) generalisation such as aspectual [go ADJ]. Specifically, it has been
argued that speakers normally take both formal and semantic properties
into account when they assess the combinability of two linguistic elements,
and it is only in the limiting case of maximally generalised semantic restric-
tions that a consideration of formal properties alone will do (Zeschel 2007).
It follows that for less generalised constructions such as the collocations
investigated in this paper, type frequencies should be counted per semanti-
cally distinct usage class rather than for a given form in aggregate: for in-
stance, for categories that could be approximated as [go ADJ-NORMAL MENTAL
PROCESSING] (cf. 2.a) and [go ADJ-ABLE TO PAY] (cf. 2.b) in separate measures rather

than for [go ADJPROPERTY] at large. Insofar as such classes have in fact been
abstracted and one is not dealing with a genuinely isolated expression, the
current approach is therefore sympathetic to Erman and War-
ren’s (2000: 41) assumption that “what we store in some cases [of slots in
‘idioms’ and ‘fixed expressions’] is a meaning rather than a specific word.”
Apart from that, the present study embraces the usage-based assumption
that speakers retain memory traces of (various aspects of) fully specific
204 Arne Zeschel

utterance tokens that they experience (i.e. exemplar representations of par-

ticular expressions), but also extract generalisations from these memorised
units (i.e. schemas) which may become entrenched routines themselves (cf.
Abbot-Smith and Tomasello 2006; Goldberg 2006, Bybee 2006). Follow-
ing Langacker (2000), such schemas can be seen as reinforced commonal-
ities between co-activated exemplars: every time a set of different units is
activated in language processing because they are all relevantly similar to a
given target, those (and only those) properties of the activated exemplars
that are shared will be jointly reinforced. Over time, the repetition of this
process may produce a stable pattern that is not identical to, but neverthe-
less completely immanent in, its individual instantiations. Put differently,
schemas are emergent generalisations over a set of relevantly similar in-
stances (including lower-level schemas) that are repeatedly co-activated in
processing. Second, clustering exemplars for similarity (i.e. the crucial
prerequisite for later schema extraction) is assumed to be an automatic by-
product of language processing: metaphorically speaking, hearers/language
learners are assumed to either map an incoming target onto an already ex-
isting identical representation in their repertoire, strengthening it, or if none
such unit is available, to store it ‘next to’ the most similar existing unit in
their system, creating a new representation (Bybee 2006). Over time, grow-
ing exposure to expressions that are in some respect similar to already en-
countered ones therefore creates progressively dense exemplar clusters,
with various local schemas/similarity classes emerging from the numerous
individual categorisations that were involved.
Applied to the question of semantic extension that is investigated in this
paper, I assume that different variants of a given construction may start out
(again metaphorically speaking) from different regions of semantic space
(due to e.g. distinct conceptual motivations). Driven by speakers’ changing
communicative needs and the demand of modelling novel expressions on
closely related established ones, such variants may then begin to expand in
different semantic directions analogically. At some point, such a cluster of
related expressions may come into contact with a hitherto distinct cluster of
other such expressions, which then permits the formation of a more inclu-
sive schema. The result is an interplay of analogical extension and subse-
quent schema extraction such that “[l]ong strings of analogical extensions
lead to discrete clusters of usage, which then license the extraction of more
abstract schemas for the construction” (Israel 1996: 220).
 Exemplars and analogy 205

3. Case study: Data and method

The present section describes how the data used in this study were ex-
tracted, postprocessed, coded, and analysed. Recall from section 1 that the
overall aim of the study is to predict the distribution of novel expressions
(of a specific semantic type) from the distribution of established expres-
sions (of the same semantic type) across the relevant regions of semantic
space (i.e. those that are covered by attested instances), and to do so on
three different levels of semantic analysis: first, on the level of all figurative
uses of the German adjective tief in attributive adjective-noun contexts at
large (subdivided into individual figurative senses; cf. 4.1); second, on the
level of the most frequent of these figurative senses (subdivided into differ-
ent conceptual mappings that give rise to this sense; cf. 4.2); and finally, on
the level of the most frequent conceptual mapping within this sense (subdi-
vided into individual semantic fields that are covered by the attested instan-
ces; cf. 4.3). On all three levels, it is predicted that the distribution of novel
adjective-noun combinations across semantic space is not random, but that
it will correlate positively with the number of established adjective-noun
combinations in different regions of the overall semantic map. Diagramma-
tically, therefore, the study progressively ‘zooms in’ on a subset of the data,
with each new analysis concentrating on just the most frequent category
(solid line circle in figure 1 overleaf) emerging from the previous classifi-
cation step.
The following pages describe the methodological procedure of the study
step by step. To begin with, 10,000 tokens of the target adjective tief in all
relevant inflectional forms were extracted from PUBLIC/COSMAS-II, a
giga corpus of written German. All concordance lines were removed in
which the retrieved form of tief did not function as an attributive adjective
which modified a noun that followed it within a span of up to three words.
Due to the high frequency of nominal Tiefe in my data, this reduced the
data set by more than 50%.
These data were then classified semantically using the lexicographic de-
scriptions of the Duden Deutsches Universalwörterbuch, a large corpus-
based monolingual dictionary of German.2 This dictionary lists seven dis-
tinct senses for tief, each of them possessing further subsenses. Simplifying
things a little, the main distinctions can be glossed and exemplified as in
(4):
206 Arne Zeschel

Figure 1. Outline: Classification steps

(4) a.‘of considerable extension (vertically) downwards, reaching far below’

ein tiefer Abgrund ‘a deep abyss’
b. ‘of considerable extension backwards, reaching far behind’
ein tiefer Wald ‘a deep forest’
c. ‘of considerable extension inwards, reaching far inside’
eine tiefe Wunde ‘a deep wound’
d. ‘temporally advanced, late’, ‘in the middle of a timespan’
bis tief in die Nacht ‘until deep in the night’
e. ‘intense, strong, given to a large extent’
tiefe Trauer ‘deep sadness’
f. ‘profound; not superficial, but reaching to the core’
eine tiefe Einsicht ‘a deep insight’
g. ‘intense, full, dark’ (of colours); ‘low, dark’ (of voices or sounds)
ein tiefes Rot ‘a deep red’
 Exemplars and analogy 207

It is not difficult to see how these senses are connected and how some of
them might have evolved from others on the list. Consequently, semantic
coding decisions were not always entirely straightforward. For instance, is
the expression tiefe Überzeugung (‘deep conviction’) an instance of sense 5
(i.e., a strong conviction) or an instance of sense 6 (i.e., a profound convic-
tion), and does it in fact makes sense to draw a line at this point? Because
the classification task is really the crucial part of the study, the data were
analysed by two separate coders and interrater agreement measures were
calculated to assess the reliability of the obtained judgments. Consistency
was assessed in two steps: First, the second coder (a linguistically naïve
native speaker of German) coded 10% of the data (424 observations), and a
first measure of interrater agreement was calculated. Second, all cases in
which there was disagreement between coders were inspected again in
order to remove differences due to oversight or carelessness, and to resolve
cases that had been explicitly coded as ‘unclear’ by one of the raters. When
these were settled, the remaining differences were used to calculate the
final agreement score. Interrater agreement was quantified using Brennan
and Prediger’s free-marginal kappa (κ, cf. Brennan and Prediger 1981), a
measure that is less likely to exaggerate the amount of actual agreement
between coders than a simple percentage because it also takes the amount
of expected chance agreement into account. As a rule of thumb, a kappa of
.7 or higher is commonly accepted as an indication of promising interrater
agreement.
In addition to sense group membership, each concordance line was
coded for whether tief was used figuratively (again by both coders, follow-
ing the procedure laid out above). Instances of senses 4-7 (cf. 4.d–g) were
coded as figurative across the board. By contrast, instances of the three
spatial senses 1-3 (cf. 4.a–c) were treated as non figurative, with two im-
portant provisos: first, the dictionary listed the meaning ‘low on a scale’ as
a subsense of spatial sense 1, ‘reaching far downwards, low’. Instances of
this subsense (e.g., tiefe Temperaturen, ‘low temperatures’) were classified
as figurative throughout. Second, judgments of figurativity were based on
assessing the meaning of the composite adjective-noun expression in its full
context (i.e. by inspecting each individual concordance line) rather than by
considering just the decontextualised noun alone. This was necessary be-
cause one and the same adjective-noun combination can be unambiguously
concrete in meaning in one context and unambiguously abstract / meta-
phorical in others:
208 Arne Zeschel

(5) a. Zwei tiefe Wunden mussten genäht werden

two deep wounds must sutured become
‘Two deep wounds had to be sutured.’
b. In die Seele der Kinder schlägt
into the soul of.the children strikes
die Obdachlosigkeit tiefe Wunden.
the homelessness deep wounds
‘Homelessness causes severe damage to children’s souls.’
To elaborate, the point is that if Wunden in (5.b) conveys the schematic
meaning ‘damage’ rather than the more specific meaning ‘laceration of the
skin’, it also coerces a fitting more abstract reading on its modifier tief (i.e.,
intensifying ‘severe’ rather than spatial ‘reaching far inwards’).
Next, the modified nouns were lemmatised: case, number and gender
markers were stripped off, spelling variants were homogenised and diminu-
tive affixes were removed. In contrast to diminutive formations, com-
pounds were left intact, to the effect that words like Krise, ‘crisis’,
Formkrise, ‘form crisis’ and Ehekrise, ‘marriage crisis’ etc. were all
counted as separate types. The resulting list was then partitioned into ‘es-
tablished’ types on the one hand and ‘novel’ types on the other by using the
simplest possible operationalisation: any combination that was recurrent
was counted as ‘established’, and any combination that occurred only once
was counted as ‘novel’. 3
Once all combinations were coded for a.) figurativity (yes/no), b.) the
particular sense of tief that they instantiated (1-7), and c.) their convention-
ality (established/novel), a possible connection between the semantic distri-
bution of established and novel types could be investigated by submitting
type counts for both classes in all five figurative sense categories to a Cor-
relation Analysis. Correlation analysis is a statistical method for assessing
dependencies between different measurements of two variables X and Y
(here: number of established vs. number of novel types in a given semantic
category). Correlations may be negative or positive in polarity, and small
(±0.1-0.3), medium (±0.3-0.5) or large (±0.5-1.0) in size. Standard meas-
ures such as Pearson’s product moment correlation coefficient (the one
used here) can be calculated by freely available statistics packages like R or
Spreadsheet software such as OpenOffice Calc. In interpreting the results of
such an analysis, it is necessary to bear in mind that the existence of, say, a
large positive correlation between two variables X and Y does not in itself
provide evidence that values for Y in fact depend on those for X: it could
just as well be the other way round (Y is the independent variable, and X’s
 Exemplars and analogy 209

value depends on Y), or both variables may be affected in similar ways by

an unknown third variable Z without there being a direct connection be-
tween X and Y. However, given the well-established connection between
type frequency and productivity, I would argue that the possibility of mix-
ing up cause and effect (or even failing to recognise an underlying ‘true’
predictor) can be neglected here. In other words, if a significant positive
correlation between the distribution of established and novel types can be
found, this will be interpreted as evidence that creative extensions of differ-
ent constructional variants are in fact driven by variants’ semantic type
frequency. The results of the first classification step and the correlation
results on this level of analysis are reported in section 4.1.
Next, the top frequent figurative sense emerging from the classification
in step 1 was subject to a more fine-grained semantic analysis. Since it
turned out that instances of the same figurative sense could be derived from
different non-figurative source senses (and connected to these through dif-
ferent conceptual mappings, cf. 4.2), it seemed useful to refine the estab-
lished classification first by systematising the data in terms of the different
conceptual mappings that gave rise to their figurative interpretation. In this
classification step, the second rater coded 50% of data (150 adjective-noun
combinations). When the classification was complete, a second Correlation
Analysis was conducted that is reported in section 4.2.
Third, one of these mappings emerged as the dominant mechanism that
accounted for the lion’s share of expressions within the most frequent figu-
rative sense. In the final classification step, all instances of this mapping
were systematised yet further by grouping them into a number of distinct
semantic fields. Again, the second rater coded 50% of the data (85 pairs) in
order to assess whether the proposed categorisation also made sense to
other native speakers, and a final Correlation Analysis was carried out on
the level of the identified semantic fields (cf. 4.3).

4. Case study: Results

The following results are based on 4245 tokens (1110 types) of attributive
tief- N expressions. Exactly two thirds of the data (66.6% of the tokens and
66.8% of the types) were identified as figurative. Interrater agreement on
this classification was only moderate in the first classification round (κinitial
classification = .58) due to an unexpected misunderstanding and following sys-
tematic mistake on part of the second coder,4 but it rose to a substantial κfinal
210 Arne Zeschel

classification=
.82 after the second round. Figurative uses were found to be dis-
tributed over 1564 established (55%) and 1266 novel types (45%).

4.1. Classification I: Figurative tief

In spite of the difficulties anticipated in section 3, interrater agreement on

sense classifications for figurative tief was already reasonably high in the
first round (κinitial classification = .77), and it rose to a kappa of .82 after a recon-
sideration of the controversial cases. The one principled disagreement that
remained (besides a number heterogeneous others) concerned the boundary
between senses 5 (‘intense, strong’) and 6 (‘profound’) that was already
mentioned in section 3. In particular, there was disagreement about the
classification of expressions from one specific semantic field, with contro-
versial examples being expressions like e.g. tiefer Glaube ‘deep faith’, tiefe
Frömmigkeit ‘deep piety’ and tiefes Bekenntnis ‘deep confession’.
From a Cognitive Semantic point of view, such difficulties are of course
anticipated since the sense distinctions provided by the dictionary do not
define clearly delimited categories with a binary membership status. Ra-
ther, one would expect members of a category to be connected by a net-
work of family resemblances, with clear instances of category X gradually
shading off into clear instances of category Y, and certain instances falling
right in between these central types. However, since only instances of the
top frequent sense 5 (‘intense’) were included in the follow-up classifica-
tions reported in sections 4.2 and 4.3, it was necessary to make a principled
decision here about what should be done with any residual disagreements.
Seeing that the second coder had only coded 10% of the data, it was de-
cided that the (final) judgments of the first coder formed the basis for all
subsequent analyses, and it is these classifications that are reported in all
following results. Table 1 on the following page shows the distribution of
recurrent and novel types across the five recognised figurative senses. The
results show that intensifying tief (e.g. tiefe Dankbarkeit, ‘deep grateful-
ness’) is both most common (i.e. token frequent) and most widely dispersed
semantically (type frequent), followed by uses conveying the meaning ‘low
on a scale’ (e.g. tiefer Zinssatz, ‘low interest rate’), followed by uses of the
type ‘profound’ (e.g. tiefes Geheimnis, ‘deep mystery’), followed by meta-
phors for particular visual and auditory qualities (e.g. tiefe Stimme, ‘deep
voice’), and, finally, a fairly rare temporal sense (e.g. tiefer Winter, ‘deep
Winter’):
 Exemplars and analogy 211

Table 1. Figurative tief: Distribution of established and novel types across senses

Rank Sense Tokens Types Est. Novel

1 5 (‘intense’) 1456 300 143 157
2 1 (‘low on a scale’) 636 198 55 143
3 6 (‘profound’) 442 137 41 96
4 7 (‘dark/low, of colour or sound’) 261 97 29 68
5 4 (‘temporal peak’) 35 9 5 4
Total 2830 741 273 468

Moving on to the productivity issue, a Correlation Analysis for the values

in the two rightmost columns in table 1 reveals the predicted positive corre-
lation between the number of established and novel types in a given sense
class (rPearson = .82). However, this result is only marginally significant at p
= .086. In other words, presumably not least because there were only five
categories/measurement pairs that were being related, this finding does not
yet provide particularly compelling evidence for the hypothesised connec-
tion between productivity and semantic type frequency. Moreover, in order
to get to the analogy issue, it will be necessary to delve deeper into the data
and examine the internal structure of the clusters that make up the individ-
ual senses, which is here demonstrated on the example of the most type-
and token-frequent class, i.e. tief as an intensifier.

4.2. Classification II: Intensifying tief

The 300 combination types that were judged as intensity expressions were
still rather heterogeneous semantically – both in terms of the meaning of
the intensified nouns and how this meaning was connected to a given non
intensifying source meaning of the adjective. Consider the examples in (5):
(6) a. tiefe Enttäuschung, tiefes Bedauern, tiefe Genugtuung
deep disappointment deep regret deep satisfaction
b. tiefe Krise, tiefe Rezession, tiefe Not
deep crisis deeprecession deep affliction
c. tiefer Graben, tiefe Kluft, tiefer Riss
deep rift deep chasm deep crevice
These figurative expressions are spin-offs from different non-metaphorical
source meanings, and connected to these meanings through different con-
212 Arne Zeschel

ceptual metaphors. In particular, the collocations with emotion and mental

state terms in (6.a) are motivated by an interplay between the metaphors
PHYSICAL AND PSYCHOLOGICAL PROPERTIES ARE ENTITIES WITHIN A PER-
SON on the one hand and CENTRALITY IS IMPORTANCE on the other: the
first metaphor allows the construal of abstract properties and sensations as
reified objects that are located within a person, and the second metaphor
contributes the intensity implication (i.e., it is ‘proximity’ to the metaphori-
cal ‘centre’ of a person that allows the construal of strong and intensely
experienced emotions as being deep-seated). Expressions of this type are
thus metaphorical spin-offs from spatial sense 3, ‘reaching far inward’
(though their ordinary use in contemporary German is arguably better char-
acterised as abstract and scalar rather than as concrete and spatial). The
expressions in (6.b) – a class of collocations denoting unfavourable situa-
tions – are rather different: motivated by a different metaphor (GOOD IS UP),
they also go back to a different spatial sense (i.e., sense 1, ‘reaching far
below’). And the expressions in (6.c) are different again: though also spin-
offs from sense 1 (‘low, below’), these collocations draw on a yet different
metaphor. Here, the mapping is provided by SIMILARITY IS CLOSENESS: the
noun denotes something that prevents two dissenting parties from moving
towards each other and eventually ending up at the same position, with the
addition of tief reinforcing the difficulty that any attempt to reconcile the
different standpoints would bring.
All in all, nine such classes could be distinguished in the data. The re-
maining six (exemplified in (7) below) can be characterised as follows:
ABSTRACT RELATIONS ARE PHYSICAL CONNECTIONS (7.a), CONTROL IS UP
(7.b) and SOUNDNESS IS PHYSICAL INTEGRITY (7.c) in terms of metaphors,
the two metonymies BOWING FOR RESPECT (7.d) and RESPIRATION FOR
RELIEF (7.e, only two types), and finally a small ‘generalised intensity’
class whose members could not be traced back to any of these original mo-
tivations but seemed to have been formed in analogy to existing tief-+N
intensity expressions (7.f):
(7) a. tiefe Bindung, tiefer Halt, tiefe Verankerung
deep bond deep footing deep anchoring
b. tiefe Bewusstlosigkeit, tiefe Narkose, tiefe Trance
deep unconsciousness deep narcosis deep trance

c. tiefe Wunde, tiefe Schramme, tiefe Verletzung

deep wound deep scratch deep injury
 Exemplars and analogy 213

d. tiefe Verbeugung, tiefer Kniefall, tiefer Kotau

deep bow deep prostration, deep kowtow
e. tiefes Aufatmen, tiefes Durchatmen
deep respiration deep respiration
f. tiefes Bouquet, tiefe Atmosphäre, tiefe Ungerechtigkeit
deep bouquet deep atmosphere deepinjustice
Interrater agreement for these classifications was lower than for sense
group membership in the first round (κinitial classification = .69). In part, this was
due to the fact that the second coder was linguistically naïve and there was
some confusion about the concept of conceptual metaphors and metony-
mies. In particular, semantic similarity in the target domain was repeatedly
favoured over isomorphism of mapping as the crucial classification cri-
terion – for instance, expressions like tiefe Ehrerbietung (‘deep reverence’),
tiefe Bewunderung (‘deep admiration’) and tiefe Hochachtung (‘deep re-
spect’) were grouped together with tiefe Verbeugung or tiefe Verneigung
(‘deep bow’) even though only the latter involve the postulated metonymy
(BOWING FOR RESPECT). When this was clarified, interrater agreement rose
to an extraordinarily high κ final classification = .98. Table 2 overleaf shows the
distribution of established and novel types across the nine attested map-
pings. A Correlation Analysis on this level of classification revealed a strik-
ingly large positive correlation between the counts for novel and estab-
lished types (rPearson = .97; p < .001***).
Table 2. Intensifying tief: Distribution of established and novel types across map-
pings

Rank Mapping Types Est. Novel

1 CENTRALITY IS IMPORTANCE 170 86 84
2 GOOD IS UP 45 17 28
3 SIMILARITY IS CLOSENESS 28 14 14
4 CONTROL IS UP 20 14 6
5 GENERALISED INTENSITY 16 15 1
6 ABSTRACT RELATIONS ARE PHYS. CONNECTIONS 10 5 5
7 METONYMY: BOWING FOR RESPECT 5 3 2
8 SOUNDNESS IS PHYSICAL INTEGRITY 4 2 2
9 METONYMY: RESPIRATION FOR RELIEF 2 1 1
Total 300 157 143
214 Arne Zeschel

4.3. Classification III: CENTRALITY IS IMPORTANCE

Again, one class made up for the lion’s share of the types in the preceding
classification: these were expressions denoting emotions and mental states
which involved a combination of the metaphors PHYSICAL AND PSY-
CHOLOGICAL PROPERTIES ARE ENTITIES WITHIN A PERSON and CENTRALITY
IS IMPORTANCE. In the final classification step, instances of this category
were systematised further by tracing them back to more or less closely cir-
cumscribed semantic fields. The 12 classes emerging from this analysis are
exemplified in (8):
(8) a. tiefe Ergriffenheit, tiefe Empfindung, tiefes Gespür
deep emotion deep sensation deep intuition
b. tiefe Freude, tiefes Glück, tiefeBefriedigung
deep joy deep happiness deep satisfaction
c. tiefe Verzweiflung, tiefe Depression, tiefe Reue
deep despair deep depression deep regret
d. tiefe Liebe, tiefe Zuneigung, tiefe Bewunderung
deep love deep affection deep admiration
e. tiefer Hass, tiefe Abneigung, tiefe Verachtung
deep hatred deep aversion deep contempt
f. tiefe Besorgnis, tiefe Angst, tiefe Beunruhigung
deep anxiety deep fear deep disturbance
g. tiefer Wunsch, tiefes Bedürfnis, tiefe Sehnsucht
deep wish deep need deep longing
h. tiefe Anteilnahme, tiefes Mitleid, tiefes Erbarmen
deep solicitousness deep compassion deep mercy
i. tiefes Unverständnis, tiefe Missachtung, tief Selbstmissachtung
deep lack of understanding deep disregard deep self-disregard
j. tiefe Verantwortung, tiefe Loyalität, tiefes Pflichtgefühl
deepresponsibility deep loyalty deep sense of duty
k. tiefe Überzeugung, tiefer Glaube, tiefes Vertrauen
deep conviction deep faith deep trust
l. tiefe Zweifel, tiefe Skepsis, tiefes Misstrauen
deep doubts deep scepticism deep distrust
 Exemplars and analogy 215

Interrater agreement for this classification was again less than perfect in the
first round (κinitial classification = .68). Here, the reason was that the analysis had
now arrived at a fairly fine-grained level of semantic description: even with
the 12 classes in (8) pre-established by the first coder prior to classification
(thus suggesting a selection of relevant semantic dimensions in advance),
there often remained more than one way of drawing the boundaries be-
tween classes in the data. On the other hand, since the respective other
coder’s judgments were often perfectly plausible, too, there was a strong
tendency to align judgments when the criteria for controversial assignments
were made explicit in the following discussion, which resulted in a mark-
edly higher kappa for the final classification (κfinal classification = .95). Taking
both tendencies into account, the comparably low initial kappa suggests
that the postulated distinctions may not provide optimal fit for the data,
whereas the high final kappa indicates that it is nonetheless possible to
converge on these categories with substantial agreement once more detailed
classification criteria are negotiated between coders. Taken together, this
suggests that the postulated groupings are presumably not stable categories
but constitute one (albeit plausible, see results of the second classification
round) possibility of carving up the semantic space covered by the attested
instances, which in reality form clouds that overlap in multiple ways and
that are not easily pinned down to a set linearly separable categories (again
not surprising from a Cognitive Semantic point of view).
Nevertheless, if the present category system is accepted as one such
possible approximation, it is remarkable to see that only 8 out of 170 types
(4.7%) could not be assigned to one of these 12 classes but had to be classi-
fied as ‘other’. Moreover, most of the postulated classes are related to oth-
ers in semantically transparent ways (notably via antonymy, but also
through frame-based shifts and other semantic relations). Even though such
intermediate-level patterning might constitute an interesting pointer to pos-
sibly recurrent ‘pathways’ of extension (cf. Zeschel 2007), this issue is
beyond the scope of the present paper. For now, suffice it to acknowledge
that the investigated collocation clusters show a remarkable degree of regu-
lar semantic patterning on different levels of schematicity. Table 3 on the
opposite page reports the distribution of established and novel types across
the 12+1 classes on this level. As in the case of the previous classification,
a Correlation Analysis revealed a large positive correlation between the
counts for novel and established types (rPearson = .78; p < .01**).
216 Arne Zeschel

5. Summary and conclusion

On all three levels of semantic analysis, large positive correlations were

found between the semantic type frequency of recurrent vari-
ants(‘established uses’) of a given expression and the number of singleton
types (‘creative extensions’) in the given cluster: even though the results of
the first (sense-level) classification were only marginally significant, they
were still pointing in the expected direction; the other two analyses pro-
duced significant results. In sum, the results thus support the hypothesis
formulated in the introduction: there does seem to be a connection between
the number of established variants of a given expression (of a given seman-
tic type) and speakers’ readiness to extend this expression to novel variants
of the same semantic type.
Table 3. Dominant mapping: distribution of established and novel types across
fields

Rank Semantic field Total Est. Novel

1 UNHAPPINESS: X feels bad 35 19 16
2 TRUST/CONVICTION: X has faith in something 18 10 8
3 STRONG IMPRESSION: X is moved by something 17 10 7
4 AFFECTION: X likes something 15 8 7
5 ANIMOSITY: X dislikes something 15 6 9
6 ANXIETY: X is worried 12 7 5
7 DISTRUST/UNCERTAINTY: X has no faith in sth. 12 7 5
8 EMPATHY: X shows concern for something 11 5 6
9 DESIRE: X wants something 10 5 5
10 HAPPINESS: X feels good 9 7 2
11 other 8 2 6
12 RESPONSIBILITY: X feels responsible for sth. 5 0 5
13 DISREGARD: X shows no regard for sth. 3 0 3
Total 170 86 84

Methodologically, the study has demonstrated how an interesting but

speculative idea from the literature can be turned into a testable hypothesis
and evaluated statistically. In addition, it has employed a procedure that
seeks to counterbalance the (inevitable) idiosyncrasy of purely introspec-
tive semantic judgments, a notorious problem for semantic classifications
5
within any framework. Even though a number of questions remain, I be-
lieve that an approach which combines careful and exhaustive semantic
 Exemplars and analogy 217

analysis of naturally occurring data with validations across different coders

(or else some sort of experimental evaluation of the obtained results) takes
Cognitive Semantics into the right direction.
In terms of linguistic model-building, the results of the study support the
usage-based assumption that speakers store linguistic units of varying sizes
(including fully predictable collocations) and re-use these representations
for later linguistic categorisations. In addition, they also support the as-
sumption that speakers sort and cluster their input using both formal and
semantic criteria, thus arguing against purely form based accounts of type
frequency and the way speakers extract productive schemas from concrete
expressions.
Put differently, the main implication of this finding for studies of pro-
ductivity in general is that for the vast majority of items that can fulfil a
particular linguistic function (here: intensification), it is actually misleading
to speak of the productivity of this element in aggregate (here: to quantify
the productivity of intensifying tief at large, as compared to other such in-
tensifying adjectives): as long as the interplay of analogy and schematisa-
tion has not yet produced a completely generalised formative (i.e. some-
thing like very in the domain of adjective intensification), calculating
overall scores will simply obscure the fact that productive use is in fact
restricted to more or less closely circumscribed semantic ‘islands’ within
the more extensive total usage spectrum of the item.

Notes

1. I thank Bettina Rust for coding part of the data, the audiences at ICLC 10,
Kraków, and DGKL/GCLA 3, Leipzig, for discussion and two anonymous re-
viewers for helpful comments.
2. While it is problematic to categorically rule out any such combination a priori,
there are nevertheless clear differences in degree of conventionalisation.
Hence, expressions marked ‘?’ do not have a single instance in the 100m word
British National Corpus, and expressions marked ‘??’ do not even yield a sin-
gle hit on the web with Google (query: went X, last checked: 30-9-08).
3. Duden: Deutsches Universalwörterbuch. Mannheim: Bibliographisches Insti-
tut, 2006. Only two observations (0.04%) could not be assigned to any of the
sense groups in (4) due to missing or unrecoverable context.
4. A reviewer remarks that an expression which occurs only once in the corpus is
not necessarily unconventional and hence creatively assembled or ‘novel’.
218 Arne Zeschel

Conversely, an expression which occurs more than once in the corpus is not
necessarily firmly ‘established’ in the speech community and hence cogni-
tively entrenched. I agree that these are certainly simplifications, but the oper-
ationalisation of creative formations as hapax legomena in a sufficiently large
corpus is neither unusual nor implausible (cf. e.g. Baayen & Lieber [1991]
and Bybee and Eddington [2006] for discussion and application of the same
strategy). Also, “more than one token in a large corpus” may not be an ideal
operationalisation of “sufficiently frequently experienced to be familiar”, but
there is no evident alternative cut-off frequency that would be less question-
able in this respect. Ultimately, the precise relationship between textual fre-
quency counts in corpora on the one hand and degrees of cognitive entrench-
ment in an idealised speaker-hearer (whose linguistic experience the corpus is
assumed to represent) remains unclear and probabilistic, not least because
speakers’ linguistic inventories are constantly changing, and no two linguistic
inventories of members of one and the same speech community are in fact
fully identical.
5. Instead of distinguishing spatial source meanings from non-spatial figurative
meanings, the second coder mistakenly drew the boundary between the
Duden’s prototypical spatial meanings (e.g. ‘reaching far (vertically) down-
ward’: ein tiefes Loch, ‘a deep hole’) on the one hand and all other mean-
ings/subsenses on the other, which also included shifted though still spatial
variants of literal meanings 1-3 (e.g. ‘low, not far from the lower boundary of
something’: tiefe Wolken, ‘deep clouds’).
6. For instance, other coders may arrive at a different category scheme to begin
with if they were not already confronted with a set of pre-established classes
by the first coder. Also, it is not clear how many categories there should be in
the first place (which is problematic since agreement will of course be higher
if there are just two or three rather general classes than if there are, say, 20
categories involving very subtle distinctions).

References

Abbot-Smith, Kirsten and Michael Tomasello

2006 Exemplar-learning and schematization in a usage-based account of
syntactic acquisition. The Linguistic Review 23: 275–290.
Baayen, Harald and Rochelle Lieber
1991 Productivity and English derivation: A corpus-based study. Linguist-
ics 29: 801–843.
 Exemplars and analogy 219

Brennan, Robert L. and Dale J. Prediger

1981 Coefficient Kappa: Some uses, misuses, and alternatives. Educa-
tional and Psychological Measurement 41: 687–699.
Bybee, Joan L.
1985 Morphology: A study of the relation between meaning and form.
Amsterdam / Philadelphia: John Benjamins.
1995 Regular morphology and the lexicon. Language and Cognitive Pro-
cesses 10: 425–455.
2006 From usage to grammar: The mind’s response to repetition. Lan-
guage 82(4): 711–733.
Bybee, Joan L. and David Eddington
2006 A usage-based approach to Spanish verbs of ‘becoming’. Language
82(2): 323–355.
Erman, Britt and Beatrice Warren
2000 The Idiom Principle and the Open Choice Principle. Text 1: 29–62.
Goldberg, Adele
2006 Constructions at Work. The nature of generalization in language.
Oxford: Oxford University Press.
Israel, Michael
1996 The Way constructions grow. In Conceptual Structure, Discourse
and Language, Adele Goldberg (ed.), 217–230. Stanford: CSLI.
Langacker, Ronald W.
2000 A dynamic usage based model. In Usage-based Models of
Language, Michael Barlow and Suzanne Kemmer (eds.), 1–63.
Stanford: CSLI Publications.
Zeschel, Arne
2007 Delexicalisation patterns: A corpus-based approach to incipient
productivity in ‘fixed expressions’. PhD dissertation, Universität
Bremen.
Section III

Advancing the science: Methodological questions

Marrying cognitive-linguistic theory
and corpus-based methods:
On the compositionality of English V NP-idioms

Stefanie Wulff

Abstract
This paper presents a corpus-linguistic compositionality measure for V NP-
Constructions (make a point, take the plunge) that implements fundamental as-
sumptions of Construction Grammar:
- the semantic contributions made by all component words to the constructional
meaning are quantified;
- the component words’ semantic contributions to the constructional meaning are
weighted as a function of their cognitive entrenchment;
- the component words’ semantic contributions are assessed both in terms of how
much of their own meaning potential they contribute to the construction, and
conversely, how much of the constructional semantics is accounted for by the
component word.
The results obtained for more than 13,000 tokens of 39 V NP-Constructions ex-
tracted from the British National Corpus tie in well with established findings from
psycho-linguistic research. Ultimately, the model strengthens the case for usage-
based approaches to grammar, demonstrating that a seemingly intuition-based,
complex phenomenon can be modeled bottom-up using performance data.
Keywords: collocation, compositionality, constructicon, Construction Grammar,
corpus linguistics, entrenchment, idioms, semantic similarity, semantic weight, V
NP-Construction

1. Introduction

One of the most popular topics in contemporary cognitive-semantic re-

search is phraseology, more specifically, the nature of phraseological ex-
pressions and their role in language theory. The growing recognition of
phraseological units as a major part of language has sparked interest in
quantitative approaches to compositionality, which appears to be a very
224 Stefanie Wulff

important parameter along which phraseological expressions vary. Compo-

sitionality is defined as “the degree to which the phrasal meaning, once
known, can be analyzed in terms of the contributions of the idiom parts”
(Nunberg et al. 1994:498). An adequate compositionality measure would
constitute an important step towards the (automatic) identification of
phraseological units, which in turn may inform the daily work of lexicogra-
phers and language teachers, further progress in fields of natural language
processing such as machine learning and machine translation, and ulti-
mately advance our theoretical understanding of the most fundamental is-
sue in linguistics, the creation of meaning.
While a number of compositionality measures have been proposed
lately, the issue I would like to address in this paper is that although these
approaches are often very impressive with regard to the quantities of data
taken into consideration and the computational skills involved, it appears
that most of these measures are not explicitly framed in any particular theo-
retical framework. Little attention is paid to the question if and to what
extent the assumptions underlying the compositionality measure are com-
patible with a theory of language at a more general level. What is more, the
quality of the data used to test the measures is often compromised in favour
of large sample sizes and extraction speed such that the constructions of
interest are extracted fully automatically and not checked manually for
false hits, which may, of course, distort the results considerably (and even
affect them adversely such that the measures proposed may indeed perform
much better in reality than they do on these unreliable data sets). This paper
presents a step towards closing these gaps by presenting a corpus-linguistic
compositionality measure for V NP-Constructions (make a point, take the
plunge) extracted from the British National Corpus; cf. section 3. The
measure presented in section 4 goes beyond previous approaches in trying
to implement fundamental assumptions of cognitive-linguistic / construc-
tionist approaches to language, more specifically, a constructionist perspec-
tive that is briefly outlined in section 2. Section 5 presents the results
thereby obtained; section 6 concludes.

2. A constructionist perspective on compositionality

Our understanding of compositionality (sometimes also referred to as de-

composability), that is, the degree to which the meanings of the component
words of a phrasal expression contribute to the meaning of that phrase, has
 The compositionality of English V NP-idioms 225

undergone dramatic changes over the last decades. For most of the last
century, the discipline of linguistics has been dominated by a Chomskyan
view of language, in which compositionality was defined as a binary con-
cept that divided language into idioms and non-idioms. Idioms seemed to
undermine the sharp division between grammar and lexis as assumed by
generative approaches: on the one hand, they behave like composite, rule-
derived structures with regard to their potential to undergo syntactic trans-
formations or modifications; on the other hand, idioms encode unitary se-
mantic concepts, which makes them word-like. Consequently, idioms were
mostly treated as exceptions and a phenomenon that is marginal to lan-
guage (Sonomura 1996).
More recently, however, the generative-transformational paradigm with
its sharp distinction between syntax and the lexicon, its primary emphasis
on syntax and relative neglect of semantics for an adequate description of
language, and its claim that the core grammar of the human language fac-
ulty is actually innate, triggered a variety of critical responses from the
fields of linguistics, psycholinguistics, and psychology. Together, these
responses have contributed to establishing Cognitive Semantics as a field of
research. Discourse-analytical approaches lay the foundation for a new role
of semantics in linguistic research by emphasizing that phraseological units
are not a marginal phenomenon in language, but on the contrary are highly
prominent and therefore indispensable units of a language (Wray 2002).
This view ultimately also entailed that the boundaries between idioms, col-
locations, and other multi-word units are fuzzy; idioms and collocations are
supposed to overlap to some extent on a continuum of fixed expressions
(Fernando 1996).
A scalar conception of compositionality also received considerable em-
pirical support from various psycholinguistic studies. All in all, these stud-
ies suggest that the literal meanings that are activated during processing
facilitate idiomatic construction comprehension to the extent that they over-
lap with the idiomatic meaning. To give but one example, Gibbs and col-
leagues (Gibbs and Nayak 1989; Gibbs et al. 1989) demonstrated that sub-
jects can distinguish between at least three classes of idiomatic
constructions in terms of their compositionality, and that sentences contain-
ing decomposable constructions are read faster than those containing non-
decomposable constructions (for further studies making similar points, cf.
Peterson and Burgess 1993; Titone and Connine 1994; McGlone et al.
1994; Glucksberg 1993).
226 Stefanie Wulff

In this paper, I adopt a constructionist perspective on compositionality,

more specifically, the Goldbergian version of Construction Grammar
(Goldberg 1995, 2006); in the following, I will briefly outline how this
framework elegantly handles the above-mentioned findings.
In Construction Grammar, a construction is defined as a symbolic unit,
that is, a form-meaning pair; given this definition, constructions are not
restricted to the level of words, but pervade all layers of language, from
morphemes to words to combinations of words. The side-by-side of sim-
plex and complex as well as lexically specified and totally unspecified con-
structions which are continuously changing on the basis of language input
and output ultimately supersedes the distinction between syntax and the
lexicon. Instead, the totality of these constructions is often assumed to be
stored in the so-called ‘constructicon’, an expanded lexicon. The construc-
ticon can basically be described as a network of constructions which is
organized in analogy to what is known about other conceptual categories,
i.e., principles like inheritance, prototypicality, and extensions are some of
the major organizational principles. Consider Table 1 for a schematic repre-
sentation of the constructicon.
Table 1. A schematic representation of the constructicon (from Goldberg 2006:5)
Morpheme pre-, -ing
Word avocado, anaconda, and
Complex word Dare-devil, shoo-in
Complex word (partially filled) [N-s] (for regular plurals)
Idiom (filled) going great guns, give the Devil his due
jog <someone’s> memory, send <some-
Idiom (partially filled)
one> to the cleaners
The Xer the Yer (e.g. the more you think
Covariational Conditional
about it, the less you understand)
Subj Obj1 Obj2 (e.g. he gave her a fish
Ditransitive (double object)
taco; he baked her a muffin)
Subj aux VPpp (PPby) (e.g. the armadillo
Passive was hit by a car)

As can be seen in Table 1, constructions range from simple and fully lexi-
cally specified to increasingly complex, lexically partially or not specified
constructions. Lexically nearly or fully specified complex constructions,
which occupy the middle range of the constructicon, are referred to as
idioms, which may be interpreted as implying that they are non-
compositional (to a considerable extent at least). However, it should be
 The compositionality of English V NP-idioms 227

noted here that while the integration of non-compositional expressions has

been a major impetus for the development of Construction Grammar, com-
plex constructions are not necessarily non-compositional: even highly
compositional expressions that are used sufficiently often to become en-
trenched in the speaker’s mental lexicon qualify as constructions (Goldberg
2006:64). Indeed, both lexico-syntactic variability and compositionality
prevail at all levels of the constructicon (Croft and Cruse 2004), albeit in
different shades of prominence and relative importance. This way, Con-
struction Grammar licenses an understanding of compositionality as a sca-
lar phenomenon that captures the whole range of constructions from fully
compositional to metaphorical to perfectly non-compositional. Accord-
ingly, the level of constructions referred to as idioms in Figure 1 can be
expected to cover constructions like write a letter (as compositional and
combinations that are assembled on the spot of words from the constructi-
con), to metaphors like break the ice, to highly non-compositional con-
structions such as kick the bucket. Only the latter are traditionally referred
to as idioms, but given our understanding of compositionality as scalar in
nature, we can regard fully compositional and fully opaque constructions as
two extremes on an idiomaticity continuum. This way, all constructions at
this level of syntactic complexity and lexical specification in the constructi-
con can be referred to as idioms, or maybe, to avoid confusion with the
traditional usage of the term, as idiomatic constructions.
Next to the conception of compositionality as a scalar phenomenon, a
constructionist perspective entails a number of working assumptions that an
adequate compositionality measure should be able to incorporate. Firstly, it
is assumed that any complex construction comprises a number of smaller
constructions, all of which make a semantic contribution to that complex
construction (in other words, constructions further up in the constructicon
as shown in Figure 1 feed into the semantics of constructions further down
of which they are part; cf. Goldberg 2006:10). In the case of V NP-
Constructions as investigated here, this means that both the verb and the
noun phrase of a V NP-Construction are expected to make a contribution.
This actually stands at odds with most existing measures, which take only
the contribution of, say, the head of a phrase into account.
Secondly, constructions are assumed to be differently entrenched in the
constructicon depending on – among other things – their frequency of use.
Accordingly, a theoretically informed measure should (i) license the possi-
bility of component words making variably large contributions (since there
is no reason to assume that verbs and noun phrases make equally large con-
228 Stefanie Wulff

tributions to V NP-Constructions), and (ii) the measure should be item-

specific in the sense that the contribution of any component word can be
differently large depending on the construction in which it occurs. For in-
stance, it is desirable to have a measure that licenses the possibility that the
contribution made by point in see a point is higher or lower than in make a
point.
In a nutshell, the agenda is to develop a corpus-linguistic measure that is
based on a large-scale data sample (in order to obtain representative re-
sults), that takes all the component words of the construction into consider-
ation, that assesses the relative contribution made by each component word
and integrates these numbers into an overall compositionality index value
for that construction, and that, last but not least, reflects potential interval-
scaled differences in the entrenchment of the complex construction and its
component lexical constructions.

3. V NP-Constructions in the British National Corpus

All instantiations of 39 V NP-Constructions (13,141 tokens total) were

retrieved from the British National Corpus (BNC), which is the largest
publicly available corpus of contemporary British English to date. The con-
struction types were primarily selected on the basis of the Collins Cobuild
Idiom Dictionary. 262 V NP-Constructions are listed in the dictionary, 33
of which occur more than 90 times in the BNC (this frequency threshold
had to be met to license statistical evaluation). While the definition of
idiom in this dictionary already captures a substantial part of the idioma-
ticity continuum, in order not to bias the sample towards one end of the
continuum, the sample was extended by another six constructions randomly
selected from a concordance of all verb-noun phrase sequences in the BNC
that occur more than 100 times (call DET police, close DET door, make
DET point, see DET point, tell DET story, and write DET letter). (1) lists
all 39 V NP-Constructions with their frequencies (in all their variant forms)
in parentheses.
(1) bear DET1 fruit (90), beg DET question (163), break DET ground (133),
break DET heart (183), call DET police (325), carry DET weight (157),
catch DET eye (491), change DET hand (212), close DET door (827),
cross DET finger (150), cross DET mind (140), deliver DET good (145),
do DET trick (155), draw DET line (310), fight DET battle (192), fill/fit
DET bill (116), follow DET suit (135), foot DET bill (109), get DET act
 The compositionality of English V NP-idioms 229

together2 (142), grit DET tooth (164), have DET clue (232), have DET
laugh (98), hold DET breath (292), leave DET mark (145), make DET
headway (136), make DET mark (213), make DET point (1,005),
make/pull DET face (371), meet DET eye (365), pave DET way (269),
play DET game (290), scratch DET head (100), see DET point (278), take
DET course (294), take DET piss (121), take DET plunge (115), take
DET root (113), tell DET story (1,942), write DET letter (1,370)

All instances of the verb and the noun phrase, both with the verb preceding
the noun phrase and vice versa, within the context window of one sentence
were searched for; the resulting concordances were cleaned manually. The
maximally inclusive search expression ensured that all matches of the con-
structions in question would be retrieved, and the manual checkup ascer-
tained that only true matches remained in the final data sample.

4. Weighting the semantic contributions of words to constructions

Starting out from the assumption that compositionality is a function of the

semantic similarity of the constituent words and the phrasal expression, a
number of corpus-linguistic compositionality measures have been proposed
lately. Some measure compositionality via the ability to replace component
words without losing the idiomatic interpretation of the construction (Lin
1999; McCarthy, Keller and Carroll 2003); others measure it via the seman-
tic similarity of the contexts of the constructions compared with those of
the component words (Schone and Jurafsky 2001; Bannard, Baldwin and
Lascarides 2003; Bannard 2005).
The measure presented here also adopts the latter approach: the working
hypothesis is that the semantic similarity of two words or constructions is
reflected in the extent to which they share collocates. Collocates of words
are “the company they keep”, that is, words that occur more often in a
(usually user-defined) context window left or right of the word than would
be predicted on the basis of the word’s general frequency. The more seman-
tically similar two words or constructions are, the more similar their con-
texts will be, which can be measured by looking into the collocates of these
words and constructions. As a matter of fact, Berry-Rogghe (1974) pre-
sented a measure along this line of thought. In a study on verb particle con-
structions (VPCs), Berry-Rogghe defined the compositionality of a VPC as
the overlap between the sets of collocates associated with the particle P and
the VPC respectively (Berry-Rogghe 1974: 21-22). Technically speaking,
230 Stefanie Wulff

this overlap can be converted into an index value R that is computed by

dividing the number of collocates shared between the VPC and the particle
by the total number of collocates of the VPC as shown in (2). R can range
between 0 when there is no overlap at all, so the VPC is perfectly non-
compositional, and 1 when the collocate sets of the particle and the VPC
match perfectly, i.e. the VPC is fully compositional.

(2)

For the present study, original R was improved in several ways. First of all,
the analysis was based on the 100 million word British National Corpus
rather than Berry-Rogghe’s 202,000 word corpus (of texts by D. Lessing,
D.H. Lawrence, and H. Fielding), which provides a more comprehensive
semantic profile of the component words and the construction.
Secondly, while Berry-Rogghe used the z-score as an association meas-
ure to identify significantly associated collocates, I opted for the Fisher
Yates Exact (FYE) test instead (which is the arguably better choice for a
variety of reasons that I cannot outline here in detail for reasons of space;
for a more detailed account, cf. Stefanowitsch and Gries 2003: 217-8).
Also, I did not include all significantly associated collocates into the collo-
cate sets that were checked for overlap; instead, collocates had to yield an
association strength of FYE≥1003 to enter into a collocate set.
Last but not least, Berry-Rogghe focused only on the contribution made
by one component word, the particle, and therefore totally disregarded any
potential contribution made by the verb. As outlined above, this procedure
is not reasonable from a constructionist perspective since it is assumed that
all component words make a contribution to the constructional meaning.
Consequently, for the V NP-Constructions investigated here, the major
question is how to combine the R-values for both the verb and the noun
phrase into one overall compositionality index value. How do we determine
their relative importance, bearing in mind the different expectations that we
have from a constructionist point of view?
The solution that I would like to suggest here is an extension of R. In a
first step, the original R-value is computed for each component word W,
which reflects how much of the semantics of a construction C that W is part
of is accounted for by W. In a second step, the original R-value is weighted
by what I will henceforth refer to as W’s share. The share reflects how
much of itself W contributes to C. Technically speaking, it is the ratio of
 The compositionality of English V NP-idioms 231

the number of collocates shared between W and C divided by the total

number of collocates of W. Consider the corresponding formula in (3).

(3) contributionW =

Let me illustrate the motivation for this approach to weighting the contribu-
tion of a component word with an example. Consider the V NP-
Construction make DET mark. Obviously, make is a high-frequency verb,
and since the number of significant collocates a word will attract is natu-
rally correlated with its overall frequency of occurrence, make has many
significant collocates (even given my highly restrictive association strength
threshold). The noun mark, on the contrary, is much less frequent, and con-
sequently, it attracts fewer significant collocates. In sum, the collocate sets
of make and mark differ in size considerably. From this, we can deduce that
make stands a much higher chance to contribute to any construction’s se-
mantics than mark; what is more, since lexically fully specified complex
constructions cannot be more frequent than their component words and
accordingly always have relatively smaller collocate sets, the resulting
overlap between a highly frequent component word’s collocates and the
construction it is part of will be quite high by default. In the case of make
DET mark, original R actually amounts to 1.0: make DET mark attracts 33
significant collocates, all of which it shares with make’s collocate set. In
other words, the semantic contribution of make to make DET mark may be
considered extremely high when looking only at how much of the construc-
tion’s semantics is accounted for. When we look into the opposite direction,
however, we see that at the same time, make contributes only a fraction of
its meaning potential: the 33 collocates it shares with make DET mark con-
stitute only a small share of its total collocate set comprising 4,234 collo-
cates. This calls for a re-evaluation of the semantic similarity between make
and make DET mark. By multiplying the original R-value by the share
value, we achieve exactly that.
For mark, a totally different picture emerges: the overlap between
mark’s collocate set and that of make DET mark is 31, which again indi-
cates a high semantic contribution, and since mark attracts 298 collocates
overall, the share of 31 out of 298 is relatively high. That is, mark is seman-
tically much more similar to make DET mark than make is in the sense that
it is much more semantically tied to this construction, while make occurs in
so many different contexts so much more often that one cannot speak of a
specific semantic association between make and make DET mark.
232 Stefanie Wulff

In sum, the overall compositionality value of a construction C is defined

as the sum of the weighted contributions of all its component words W (in
the case of V NP-Constructions, the verb and the noun phrase). In order for
the overall compositionality value to range between 0 and 1, it would prin-
cipally be necessary to divide it by the number of component words that
entered the computation (here: two); however, for the V NP-Constructions
and the VPCs discussed here, the values were extremely small already, so
the values reported here are not divided. However, this not a problem as
long as only the results of one analysis are compared (since the ranking of
the constructions remains the same), but once results from several analyses
are compared, it may be reasonable to do the division in order to stay
within the range from 0 to 1.

5. Results

Let us turn to the results for the whole data sample of V NP-Constructions.
Figure 2 provides an overview (cf. also Table 2 in the appendix for the
exact values).

Figure 1. V NP-Constructions and their weighted R-values

 The compositionality of English V NP-idioms 233

As can be seen in Figure 1, the weighted R-measure neatly reproduces what

we would expect from established idiom taxonomies (e.g. Fernando 1996;
Cacciari and Glucksberg 1991): core idioms like make DET headway and
take DET plunge rank lowest in compositionality; metaphorical expressions
like make DET mark and meet DET eye occupy the middle ranks; quasi-
metaphorical constructions, the literal referent of which is itself an instance
of the idiomatic meaning (Cacciari and Glucksberg 1991), like cross DET
finger, hold DET breath, and scratch DET head, tend to rank even higher
in compositionality; and most of the constructions that were not picked
from the idiom dictionary rank highest, with write DET letter yielding the
highest weighted R-value.
Note also that the majority of items is assigned a fairly non-
compositional value on the scale from 0 to 1, which ties in nicely with the
fact that most of these were actually obtained from an idiom dictionary;
items such as write X letter and tell X story, on the other hand, were selec-
ted so to test if items that are intuitively assessed as (nearly perfectly) com-
positional are actually treated accordingly by the measure – so weighted R
proves very accurate since these items do not only rank highest, but more-
over, their compositionality values are very high in absolute terms (.73 for
tell X story and .84 for write X letter).4

6. Conclusion

The extension of Berry-Rogghe’s R-value presented here provides very

satisfactory results, and there are several strong indications that it is indeed
valid. For one, the correlation between the weighted R-values and the cor-
pus frequency of the V NP-Constructions is very high (rPearson=.802), which
stands in accord with Barkema's (1994: 26) results, where corpus frequen-
cies were correlated with intuitively assessed compositionality values. At
the same time, the measure does not merely equate compositionality with
frequency of occurrence but also manages to model aspects of composi-
tionality that go beyond frequency, as is evidenced by the fact that the re-
sulting ranking departs from a purely frequency-based ranking for V NP-
Constructions that are relatively frequent and relatively opaque (consider,
e.g., make DET face, which is the sixth most frequent construction in the
sample with a corpus frequency of 371, yet obtains the fourth lowest R-
value of .021), and vice versa (an example here is break DET heart with
only 183 attestations in the BNC but a fairly high R-value of .238).
234 Stefanie Wulff

What is more, the corpus-linguistic definition of compositionality pre-

sented here derives a lot of plausibility from its compatibility with theoreti-
cal premises:
- the measure stands in accord with the constructionist view that a
complex construction is a manifestation of several smaller construc-
tions, and that every one of them contributes to the meaning of the
complex construction;
- the measure implements the central assumption of many cognitive
approaches to grammar that constructions are entrenched in the men-
tal lexicon to different extents;
- the measure leaves room for a potential backward influence of the
construction’s semantics on the (weightings in the often polysemous
network of) the constituent word’s semantics.
To conclude, this paper presents a first step towards a theoretically in-
formed and performance-based compositionality measure referred to as
weighted R that may be used not only for assessing the compositionality of
V NP-Constructions, but that is principally applicable to any kind of con-
struction.5 Beyond that, weighted R may prove a useful tool for quantitative
approaches to related issues in Cognitive Semantics. For instance, it could
be employed to quantify the degree of semantic bleaching of verbs, which
could inform the investigation of incipient grammaticalization processes,
particularly with regard to the question to what extent the degree of bleach-
ing is context-dependent. On this note, I hope that the present paper contri-
butes to the growing awareness of the vast potential that resides in marry-
ing cognitive-linguistic theory and corpus-based methods.6
 The compositionality of English V NP-idioms 235

Appendix

Table 2. V NP-Constructions and their weighted R-values

V NP-Construction weighted R V NP-Construction weighted R
make DET headway .003 carry DET weight .137
take DET plunge .004 follow DET suit .147
take DET piss .008 beg DET question .150
make/pull DET face .021 bear DET fruit .160
get DET act together .026 deliver DET good .161
pave DET way .033 cross DET finger .171
change DET hand .051 draw DET line .174
take DET course .058 take DET root .185
foot DET bill .058 cross DET mind .225
see DET point .062 hold DET breath .232
leave DET mark .074 break DET heart .238
grit DET tooth .079 fight DET battle .288
break DET ground .079 do DET trick .340
meet DET eye .101 make DET point .359
make DET mark .106 scratch DET head .368
have DET laugh .106 close DET door .421
fill/fit DET bill .108 catch DET eye .432
have DET clue .117 tell DET story .730
call DET police .117 write DET letter .844
play DET game .132

Notes

1. DET stands for any kind of determiner, including a zero determiner.

2. While this is not a V NP-Construction, it was included in several pre-tests and
is reported alongside the V N P-constructions.
3. The regular output of a FYE test is a p-value; since these values are sometimes
so extremely small that they become extremely difficult to interpret and cum-
bersome to report, Gries at al. (2005: 648) suggest to report the p-value’s
negative logarithm to the base of 10 instead; a converted value of ≥1.3 is equi-
valent to a 5% probability of error. Accordingly, a FYE value of ≥100 indi-
cates an extremely high association strength. A series of pre-tests to the results
reported here showed that not only is this threshold value maximally conserva-
tive with regard to which collocates enter into the collocates sets, the resulting
R-values also spread widest with this threshold, which is desirable if the whole
span of a continuum is intended to be modeled.
236 Stefanie Wulff

4. While individual rankings may not make too much sense intuitively, such as
call X police obtaining a middle rank only, this does not necessarily mean that
the measure per se is flawed; more likely, this result is a side-effect of the yet
limited sample size for these kinds of expressions.
5. Future research should systematically explore what kind(s) of measure(s) are
suited best to model the association strength between constructions at different
levels of schematization: weighted R is just one of many collocation-based as-
sociation measures that are compatible with constructionist premises, if only
highlighting different aspects of the syntagmatic and paradigmatic dimensions
of constructional interaction. A first example of such a contrastive analysis is
presented by Speelman et al. (2009), who contrast Gries and Stefanowitsch’s
collostructions (Stefanowitsch and Gries 2003) and co-varying collexemes
(Gries and Stefanowitsch 2004) for measuring the association strength be-
tween different inflectional variants of Dutch attributive adjectives and their
head nouns.
6. For a corpus-linguistic approach that addresses the question how the present
compositionality measure competes with other variables that jointly charac-
terize the overall idiomaticity of V NP-phrases, cf. Wulff (2008).

References

Bannard, Colin
2005 Learning about the meaning of verb particle constructions from
corpora. Journal of Computer Speech and Language 19(4): 467-478.
Bannard, Colin, Timothy Baldwin and Alex Lascarides
2003 A statistical approach to the semantics of verb-particles. Proceed-
ings of the ACL-Workshop on Multiword Expressions: Analysis, ac-
quisition, and treatment, 65-72.
Barkema, Henk
1994 The idiomatic, syntactic and collocational characteristics of received
NPs: Some basic statistics. Hermes Journal of Linguistics 13: 19-40.
Berry-Rogghe, Godelieve L. M.
1974 Automatic identification of phrasal verbs. In Computers in the Hu-
manities, John Lawrence Mitchell (ed.), 16-26. Edinburgh: Edin-
burgh University Press.
Cacciari, Cristina and Samuel Glucksberg
1991 Understanding idiomatic expressions: The contribution of word
meanings. In Understanding Word and Sentence, Geoffrey Simpson
(ed.), 217-240. The Hague: North Holland.
 The compositionality of English V NP-idioms 237

Croft, William and D. Alan Cruse

2004 Cognitive Linguistics. Cambridge: Cambridge University Press.
Fernando, Chitra
1996 Idioms and Idiomaticity. Oxford: Oxford University Press.
Gibbs, Raymond W.
1995 Idiomaticity and human cognition. In Idioms: Structural and Psy-
chological Perspectives, Markus Everaert, Erik-Jan van der Linden,
André Schenk and Rob Schreuder (eds.), 97-116. Hillsdale, NJ:
Lawrence Erlbaum.
Gibbs, Raymond W. and Nandini Nayak
1989 Psycholinguistic studies on the syntactic behavior of idioms. Cogni-
tive Psychology 21(1): 100-138.
Gibbs, Raymond W., Nandini Nayak, John Bolton and Melissa Keppel
1989 Speakers’ assumptions about the lexical flexibility of idioms. Mem-
ory and Cognition 17(1): 58-68.
Glucksberg, Samuel
1993 Idiom meaning and allusional content. In Idioms: Processing, struc-
ture, and interpretation, Cristina Cacciari and Patrizia Tabossi
(eds.), 3-26. Hillsdale, NJ: Lawrence Erlbaum.
Goldberg, Adele E.
1995 Constructions: A Construction Grammar approach to argument
structure. Chicago: Chicago University Press.
2006 Constructions at Work: The nature of generalization in language.
Oxford: Oxford University Press.
Gries, Stefan Th. and Anatol Stefanowitsch
2004 Co-varying collexemes in the into-causative. In Language, Culture,
and Mind, Michel Achard and Suzanne Kemmer (eds.), 225-236.
Stanford: CSLI.
Gries, Stefan Th., Beate Hampe and Doris Schönefeld
2005 Converging evidence: Bringing together experimental and corpus
data on the association of verbs and constructions. Cognitive Lin-
guistics 16(4): 635-76.
Langacker, Ronald W.
1987 Foundations of Cognitive Grammar. Vol. 1. Theoretical prerequi-
sites. Stanford: Stanford University Press.
Lin, Dekang
1999 Automatic identification of noncompositional phrases. Proceedings
of the 37th Annual Meeting of the ACL, College Park, USA, 317-
324.
238 Stefanie Wulff

McCarthy, Diana, Bill Keller and John Carroll

2003 Detecting a continuum of compositionality in phrasal verbs. Pro-
ceedings of the ACL-Workshop on Multiword Expressions: Analysis,
acquisition, and treatment, 73-80.
McGlone, Matthew S., Samuel Glucksberg and Cristina Cacciari
1994 Semantic productivity and idiom comprehension. Discourse Pro-
cesses 17(2): 176-190.
Nunberg, Geoffrey, Ivan A. Sag and Thomas Wasow
1994 Idioms. Language 70(3): 491-538.
Peterson, Robert R. and Curt Burgess
1993 Syntactic and semantic processing during idiom comprehension:
neurolinguistic and psycholinguistic dissociations. In Idioms: Pro-
cessing, structure, and interpretation, Cristina Cacciari and Patrizia
Tabossi (eds.), 201-225. Hillsdale, NJ: Erlbaum.
Schone, Patrick and Dan Jurafsky
2001 Is knowledge-free induction of multiword unit dictionary headwords
a solved problem? Proceedings of the 6th Conference on Empirical
Methods in Natural Language Processing, 100-108.
Sonomura, Marion O.
1996 Idiomaticity in the Basic Writing of American English. Formulas
and idioms in the writing of multilingual and Creole-speaking com-
munity college students in Hawaii. New York: Peter Lang.
Speelman, Dirk, José Trummers and Dirk Geeraerts
2009 Lexical patterning in a Construction Grammar: the effect of lexical
co-occurrence patternson the inflectional variation in Dutch attribu-
tive adjectives. Constructions and Frames 1(1): 87-118.
Stefanowitsch, Anatol and Stefan Th. Gries
2003 Collostructions: Investigating the interaction of words and construc-
tions. International Journal of Corpus Linguistics 8(2): 209-243.
Titone, Debra A. and Cynthia M. Connine
1994 The comprehension of idiomatic expressions: Effects of predicta-
bility and literality. Journal of Experimental Psychology: Learning,
memory and cognition 20(5): 1126-1138.
Wray, Alison
2002 Formulaic Language and the Lexicon. Cambridge: Cambridge Uni-
versity Press.
Wulff, Stefanie
2008 Rethinking Idiomaticity: A usage-based approach. London / New
York: Continuum Press.
Testing the hypothesis
Objectivity and verification in usage-based
Cognitive Semantics

Dylan Glynn

Abstract
This study has two aims: to show the methodological possibility of doing purely
subjective semantic research quantitatively and to demonstrate theoretically that
discreet senses and discreet linguistic forms do not exist. On the methodological
front, it argues that, with due caution and statistical modelling, subjective semantic
characteristics, such as affect and cause, can be successfully employed in corpus-
driven research. The theoretical implications show that we cannot treat lexical
senses as discreet categories and that the semasiological - onomasiological and
polysemy - synonymy distinctions are not tenable and must be replaced with a
more multidimensional and variable conception of semantic structure. The case
study examines a sample of 650 occurrences of the lexeme bother in British and
American English. The occurrences are manually analysed for a range of formal
and semantic features. The exploratory multivariate technique Correspondence
Analysis is used to indentify three basic senses relative to formal variation and
subjective usage-features. Two of these sense clusters are then verified using Lo-
gistic Regression Analysis. The analysis demonstrates a statistically significant
difference between the two senses and indentifies which of the semantic features
are most important in distinguishing the uses. The statistical model is powerful and
its predictive strength serves as further verification of the accuracy of the semantic
analysis.
Keywords: polysemy, semantics, objectivity, corpus linguistics, behavioural pro-
file, Logistic Regression Analysis, Correspondence Analysis

1. Introduction. Quantitative Cognitive Semantics

1
Is purely semantic research at all possible using quantitative techniques? Within
Cognitive Semantics, a range of studies have shown how a combination of observ-
able formal characteristics and semantic, yet objectively determinable, characterist-
ics yields coherent and verifiable descriptions of semantic structure (Geeraerts et
al. 1994, Gries 2006, Divjak 2006 inter alia). However, not all linguistic forms
240 Dylan Glynn

possess objectifiable semantic usage-features. Can we overcome this limitation by

examining purely semantic usage-features?
As corpus-driven research comes to the study of semantics, it inevitably
meets the question of subjectivity versus objectivity. There is an unfortu-
nate tendency to believe that corpus-driven research and the quantitative
treatment of the results obtained from such methods are inherently more
objective than traditional methods, such as elicitation and introspection.
This is not necessarily the case. The choices involved in the annotation of
data remain largely subjective. This is especially true for the study of se-
mantics. Whether approached with hermeneutics or the latest statistical
modelling, meaning is a symbolic relation created in our minds and is
therefore always beyond the reach of pure objectivity. Indeed, the ‘object’
of study does not exist, save as a subjective experience. Even the results
from a formal operationalisation of semantics (‘indirect’ methods, such as
collocation analysis) must be interpreted in subjective terms if they are to
be used to capture meaning.
Indeed, in a usage-based framework, even ‘direct’ semantic analysis of
found examples is an indirect line of inquiry: we make generalisations
about usage and then make an assumption that these usage patterns repre-
sent speaker knowledge of the symbolic relations that constitute ‘meaning’.
It is for this reason that Talmy (2007, 2008) argues that some research
questions are better answered by introspective methods. On the other hand,
Geeraerts (2007; 2010; this vol.) argues that the results obtained through
subjective experience represent merely the first step in the analytical cycle.
They are crucial to research but represent hypotheses and, as hypotheses,
need to be tested. From this perspective, the challenge is to find ways to
operationalise this infamously slippery object of study, semantics.
Within Cognitive Linguistics, this challenge is being met on a variety of
fronts. Two schools, a variationist, multifactorial approach (Heylen 2005;
Tummers et al. 2005; Divjak 2006; Grondelaers et al. 2007; Glynn 2009;
Speelman & Geeraerts 2010) and a collocation-based approach (Ste-
fanowitsch & Gries 2003; Gries & Stefanowitsch 2004; Wulff et al. 2007;
Hilpert 2006, 2009) lead the search for a response to this challenge. A third
cognitively compatible approach is emerging that employs computational
techniques. Building on the principle of collocation, it uses the word space
modelling to investigate synonymy and polysemy (Peirsman & al. 2010).
All these approaches adopt a usage-based model of language and as-
sume that patterns of usage found in corpora are indices of the grammar in
the minds of speakers in a language community. This study belongs to the
 Objectivity and verification in usage-based Cognitive Semantics 241

first of these approaches, which is specifically concerned with the multidi-

mensionality of language. In order to capture the interaction, or integration,
of lexis, syntax, morphology, prosody, and then interpret all of these factors
relative to the full gamut of sociolinguistic complexity, any analysis must
employ multivariate statistics.
Cognitive Linguistics adheres to a model of language that is semanti-
cally driven. It is argued that all form, and variation in that form, is moti-
vated by variation in meaning. Therefore, identifying patterns in form indi-
rectly identifies patterns of meaning. However, though all formal structure
is motivated by semantic structure, this does not entail that all semantic
structure is mirrored formally. Language is the vehicle for expressing
meaning and so we can assume that speakers will constantly search for
formal ways of encoding what they wish to express. Nevertheless, no one
would argue that every possible concept should be distinguished formally.
Therefore, we must assume that there is a great deal of semantic content
that cannot be formally identified. Though formal variation is a result of
semantic variation, and therefore indicative of it, it can only ever be indica-
tive of a small part of semantic structure.
The variationist and multifactorial research mentioned above restricts it-
self to relatively objective semantic features, such as the animacy or con-
creteness of the actors in an event (Gries 2006; Divjak 2006; Grondelaers
& Speelman 2007). These studies have shown how combining the analysis
of purely formal characteristics and objectifiable semantic characteristics
can accurately map the synonymy and polysemy of both grammatical and
lexical semantics. However, in all these cases, the forms in question pos-
sess rich argument structures, where relatively objective semantic charac-
teristics are available. However, not all objects of study possess such easily
identifiable features. Can we extend quantitative methods to the study of
semantic structures that are not so easily identifiable?

2. Objectivity, verification, and prediction

Sandra & Rice (1995) unequivocally demonstrated that the lexical network
analyses prevalent at the beginning of Cognitive Semantics produced ad
hoc and unverifiable results. Despite their resounding refutation, there is
nothing inherently wrong with lexical network studies of semantic struc-
ture, such as Lakoff’s (1987) analysis of over. This research, however, only
represents the first step in semantic analysis – the proposal of a coding
242 Dylan Glynn

schema. This is a crucial operationalisation for quantitative semantic analy-

sis of natural language. Operationalisation enables quantification; quantifi-
cation enables verification and thus the testing of hypotheses (see Geeraerts
this vol. and Stefanowitsch this vol.).
It is often argued, and with some reason, that subjective semantic char-
acteristics cannot be operationalised in a sufficiently objective manner to
include them in quantitative studies. There exists, however, one strong
counterargument. It is not objectivity that quantitative analysis offers us,
but a better and more varied way of verifying the results. Seen from this
perspective, quantitative methods are all the more important for subjective
semantic analysis. Since the goal of descriptive research is accuracy, when
an object of study does not possess clear observable features, verification is
essential.
There are three basic ways in which the quantitative treatment of found
data allows us to verify the results of analysis: overt operationalisation,
repeat analysis, and statistical modelling. Let us consider the importance of
each of these steps for corpus-driven quantitative analysis in semantic re-
search.

Overt operationalisation - subjective analysis

Operationalisation is crucial in quantitative research. It requires overt iden-
tification of criteria for the analysis. Without entering into a philosophical
debate, meaning exists entirely in our minds and not in the observable ‘real
world’. Thus, with no objectively observable characteristics, one cannot
operationalise direct semantic analysis. It is for this reason that much quan-
titative research has avoided semantics. Although objectively operationalis-
ing subjective features may be technically impossible, using the principle of
operationalisation to help overtly identify the analytical criteria acts as a
perfect goal for improving semantic analysis. Most importantly, it permits
quantification and therefore verification. If the semantic criteria, even sub-
jective ones, are overtly identified, a second researcher can re-apply the
analysis. Discrepancy between the analyses aids the process of refining the
operationalisation and, thus, enhances its accuracy.

Repeat analysis - multiple datasets

Even if based on subjective analysis, the two basic advantages of empirical
research are the ability to repeat the analysis on the same data and the
possibility to apply the analysis to a second dataset. The former verifies the
accuracy of the analysis, the latter verifies the representativity of the data.
 Objectivity and verification in usage-based Cognitive Semantics 243

Duplicating subjective analysis and comparing the results of different cod-

ers, or ‘raters’, has an established tradition in psychology. There is no rea-
son why this should not be applied to semantic research. Szmrecsanyi
(2003), Diessel (2008), and Zeschel (this vol.) use this technique in their
analyses and employ Cohen’s Kappa Coefficient to determine the degree of
inter-coder agreement.
The second advantage is one of the basic and unquestionable assets of
empirical research - if the criteria of the analysis are adequately operation-
alised, then the analysis can be simply repeated upon a second sample.
Although no one will argue that any sample is perfectly representative of
the population (in our case the collective utterances of an entire language),
this will give concrete information on just how representative the sample is.

Statistical modelling - quantitative evaluation

There are two basic kinds of statistical analysis: exploratory and confirma-
tory. Both types offer means for verification. Firstly, exploratory statistical
analysis identifies patterns in the data. Coherent patterns, as opposed to
random dispersion in the data, are the first verification of the analysis. If, at
this exploratory level, intuitively sound results are obtained, especially if
these results match previous introspection-derived results, we have a strong
argument that the semantic analysis was accurate.
Yet, it is with confirmatory statistics that the real power of verification
comes to the fore. Within the recent linguistic literature, there has been a
growing tendency to use quantitative data and to employ tests for statistical
significance. However, statistical modelling offers another equally, or per-
haps, more important tool - that of explanatory power. Whereas, statistical
significance merely tells us that the finding is more than chance, accuracy
scores of explanatory power are obtained through confirmatory modelling,
such as Logistic Regression Analysis. Obviously, with subjective non-
observable criteria, confirmatory modelling is a vital part of result verifica-
tion.

3. Bother. A quantitative semantic study

One of the problems for quantified semantic analysis is that beyond the
finite verb and its argument structure, it becomes difficult to operationalise
semantic features. Verbs are associated with actors and express relations
between actors. These alone offer a wide range of readily quantifiable se-
244 Dylan Glynn

mantic characteristics, such as the animacy of the actor or the boundedness

of the ground and so forth. Such semantic information, one may quite
easily operationalise and manually annotate in a sample of natural lan-
guage. Prepositions, adverbs, and adjectives are relational and, therefore,
also offer some room for this kind of indirect operationalisation of semantic
analysis. However, what of very abstract verbs such as love or hate, or ab-
stract nouns, or even discourse markers? How can one approach such lex-
emes? The lexical semantic study of emotion concepts, such as Kövecses
(1986) and Lakoff (1987), is representative of the birth of Cognitive Lin-
guistics. Yet, quarter of a century after Kövecses and Lakoff, as we apply
powerful multivariate statistics to our corpus-driven analyses, we have no
method for even beginning to approach abstract concepts like PRIDE, LOVE,
or ANGER. Does this mean that corpus linguistics is ill-suited to the study of
such lexical structure? The task of any scientist is to develop a method that
will match the object of inquiry. We can say that abstract lexical concepts
are beyond scientific reach no more than an astronomer can say the stars
are too far to be studied. This study seeks to take the first steps towards the
quantified study of abstract semantics with no identifiable referent to act as
a tertium comparationis.
The study treats the lexical concept ‘bother’, which is profiled by a noun
and a verb. It experiments with semantic analysis and the statistical treat-
ment of the results thereof to ascertain the possibility of performing purely
subjective quantified analysis. The lexeme bother is chosen because it is
sufficiently abstract to test the method, while still possessing some points
of reference, such as the Agentive cause of the emotion and the Affect of
the event upon the patient. The study tests the efficacy of different degrees
of subjectivity. Secondary annotation is also used to test the accuracy of the
analysis.

3.1. Data and analysis

The data are taken from a corpus of on-line personal diaries, or ‘blogs’, in
2
both British and American English . Over 300 examples were randomly
extracted from each of the two dialects, in total, some 650 occurrences. The
sample included no instances of bothersome, botheration, or botherly, and
only relatively small numbers of nominal and gerundive profilings. The
majority of the occurrences belonged to a range of verbal constructions that
are discussed below.
 Objectivity and verification in usage-based Cognitive Semantics 245

In order to limit the number of variables, the study is restricted to the

extralinguistic variables dialect and thematic topic of discourse. One of the
greatest limitations of any manual analysis is the small number of occur-
rences that can be treated. The labour-intensive process of manual annota-
tion restricts the sample size considerably, which, in turn, significantly
restricts the number of different factors that can be examined. The two re-
gions are broadly defined as British and American. The thematic topic of
discourse is highly subjective and is thus described below with the semantic
factors. Other possible extralinguistic factors are controlled for by the ho-
mogenous nature of the corpus. Although impossible to determine scientifi-
cally, on-line personal diaries tend to be authored by women and we can
suppose that, as a genre, they are restricted to the younger generations.

3.1.1. Objectifiable variables

The formal analysis covers the obvious grammatical features of form and
part of speech, which, for nouns, include count singular, count plural, and
mass and, for verbs, tense, aspect, mood, voice, polarity, person, and num-
ber. The frame semantics, argument structure, and syntax of the lexeme are
quite rich. There are two basic frames: the first, where the Subject co-
incides with the Agent and affects the Patient, either directly or instrumen-
tally, and the second, where the Subject is at once the Agent and the Patient
of the BOTHER event. The first of these argument structures we will term
the Agentive Frame and the second, the Predicative Frame. To understand
the difference between these two argument structures, it can be useful to
differentiate Cause and Agent. Expressions such as I bother to eat are se-
mantically a kind of reflexive. The Cause of BOTHER is ‘eating’, but both
the Agent and the Patient are the Subject of the sentence. The Subject
chooses to eat and this choice leads to the experience of BOTHER. There-
fore, ‘eating’ is the Cause, but it is not the Agent. Differences of this kind
in profiling the relationship between the emotion, its experiencer, and the
Cause of the experience vary greatly. Indeed, as we will see below, one of
the constructions associated with this second frame completely back-
grounds the Cause with no slot for its specification. Wierzbicka (1995)
offers some discussion on these kind of phenomena.
At a relatively coarse-grained level of analysis, there exist five verb-
based constructions. There were only three occurrences of resultatives,
246 Dylan Glynn

which due to their marked argument structure and rarity will not be in-
cluded in the analysis. All examples are taken from the corpus.
Direct Transitive
Subj. V Direct Object
I BOTHER you
Agent/Cause verb Patient
(1) but oh well it doesnt bother me too much because every1 is entitled to
their own opinion.

Oblique Transitive
Subj. V Direct Obj. Indirect Object
I BOTHER you with/about -over -because of it
Agent Verb Patient Cause
(2) The day carried on as boring and reptative as ever so I wont bother you
wiith the rest.

Oblique Reflexive Transitive

Subj. V Indirect Object
I BOTHER about/ over/ with it
Agent/Patient Verb Cause
(3) I will no longer bother with questions of the capabilities of people not
caring for each other. (Instrumental)

Complement Clause Transitive

Subj. Verb Infinitive / Gerund / Dep. Clause
I BOTHER to eat / eating / that I’m not eating
Agent/Patient Verb Cause
(4) a. Actually I did and no one bothers to move. (Infinitive)
b. That’s if i can be bothered goin into business! (Gerund)
c. and I was bothered that I never knew about it until just then. (Dep.
Clause)

Predicative
S COP Participle
I am bothered
Agent/Patient Verb
(5) a. but its so crusty and I really cannot be bothered :(
b. I been listening to the K’s choice and eisley CD a lot if anyones both-
ered they should download some of their music or something.
 Objectivity and verification in usage-based Cognitive Semantics 247

The final construction poses difficulties because, at first glance, it appears

to be a middle voice and therefore an elided transitive construction. How-
ever, there is no possibility of rephrasing the sentence as an active form.
This suggests we could term it an Intransitive Construction. Moreover, it
behaves, at many levels, like a Predicative Adjective. It is felicitous with
verbs other than be, for example, she seems bothered. In order to avoid
terminology debates, we will use the term Predicative Construction.
The sample also includes count singular, count plural, and mass nouns
as well as gerunds. However, they will be excluded from the analysis due to
the low number of occurrences.

3.1.2. Non-objectifiable variables

The subjective semantic analysis is the central point of the study and so
warrants a close explanation. Following previous multifactorial lexical
semantic research (Gries 2006, Divjak 2006, Glynn 2009), the actors in the
event are coded for semantic features such as animacy and specificity. We
can call this the coarse-grained Actor Semantics. However, in this study,
progressively more subjective levels of analysis are applied. In particular,
the Actor Semantics was analysed for a fine-grained, and therefore subjec-
tive, level of analysis. The features are listed below.

Cause and Agent

The semantics of the Cause and the Agent can be determined through sort-
ing and re-labelling the analysis of the Subject, Object, and Oblique. There-
fore, in order to render the subjective analysis more overt, the formal cate-
gories are analysed, not the semantic categories. Cause and Agent are
determined through post-analysis sorting. This gives us two sets of seman-
tic analysis: one linked to the formal Actors - Subject, Object, and Oblique,
and the other, to the semantic Roles - Agent and Cause. Since the Patient is
almost exclusively a specified known human, this factor is uninformative.
A fine-grained list of features for the Actors and Roles is presented below.
They are broadly grouped into things and events, though one of the fea-
tures, ‘abstract state of affairs’, represents a blurred line between the two.
This feature was typically encoded by it or this or some conceptual shell
noun (cf. Schmid 2000), but the point of reference is so abstract that it often
is actually an event of some sort.
248 Dylan Glynn

Human Specified A specific individual, or individuals, known to the

speaker
Human Non-specified Human, but a generic human unknown to the
speaker
Thing Concrete Something at which you can point
Thing Abstract A thing at which you cannot point
State of Affairs Abstr. Abstract information about the world known by the
interlocutors
State Concrete Abstract event that is profiled as durative
Event An event with a perfective profiling
Activity An event with an imperfective profiling
Up until this point, the present semantic analysis differs from the analyses
of Gries, Divjak, and Glynn only in a matter of degree of semantic granu-
larity. The fine-grained categories used to annotate the Actors and Roles
are difficult to operationalise only because of their detail. In principle, such
semantic characteristics can be quite objectively determined. However, two
other semantic analyses of the examples are performed – the Patient’s Af-
fect (or the emotion experienced) and the Thematic Subject (or the topic of
discourse).

Affect
For the Affect, the analyst must read and attempt to ascertain what kind of
emotion is being expressed by the use of the term. The only way of oper-
ationalising this variable was by composing questions and, after close read-
ing of the text, asking which of the questions most accurately captured the
emotional state of the patient. The Affect features were identified with the
questions listed below (note that for negative sentences, phrasing the ques-
tion with the hypothetical would made the analysis clearer):
Anger Does the Patient feel angry because of the event?
Anxiety Is the Patient concerned or worried by the event?
Boredom Does the Patient feel bored because of the event?
Energy Does the Patient feel tired because of the event?
Imposition Does the Patient feel imposed upon by the event? Does the
Patient have to do something she/he does not want to do?
Interruption Does the Patient feel interrupted? Is the Patient prevented
from doing something she/he wants to do?
Pain Does the Patient feel seriously upset by the event?
 Objectivity and verification in usage-based Cognitive Semantics 249

Theme
The last subjective variable under consideration is more of an extralin-
guistic factor than a semantic one. In this, it can be an important operation-
alisation of register and, therefore, representativity of a sample. It can also,
indirectly, give semantic information, since we can suppose that more seri-
ous topics of discourse will be associated with certain meanings of a word
and so forth. In general, it can be a useful factor, but it suffers from the
constraint that with small samples, keyword indices are not possible, leav-
ing only the possibility of highly subjective manual coding. In this study,
the basic distinction is between the personal sphere and society, in the
broad sense. This is then re-analysed for more fine-grained distinctions.
The personal features are by far the more common and important, due to
the genre of diary. The features include:
Society-
Entertainment Soc. Discussion about cinema, music, and sport at the
level of society
Miscellaneous Soc. Discussion at a social level not fitting into other
categories
Religion & Politics Discussion about religion, history, politics and
economics at the level of society
Personal-
Cyber Friends Discussion about friends in the cyber community
of the on-line diaries. Due to the genre, an import-
ant category
Entertainment Pers. Discussion about going out, parties concerts and so
forth
Family Discussion about family and personal relationships
Health Discussion about personal health
Miscellaneous Pers. Discussion at a personal level not fitting into the
other categories
Study & work Discussion about school, university, and work at a
personal level
Computing This refers to problems with a personal computer, a
common topic in the on-line diaries
250 Dylan Glynn

3.1.3. Inter-coder agreement

Following Szmrecsanyi (2003) and Zeschel (this vol.), secondary annota-

tion was employed. Each of these four highly subjective variables, the se-
mantics of the Agent, the Cause, the Affect, and the Theme of Discourse,
were coded independently by two linguists. After discussion and trial sam-
ples, inter-coder agreement was excellent. Cohen’s Kappa is used to calcu-
late the degree of inter-coder agreement. The rules of thumb over agree-
ment rating follow: Strength of agreement < 0.2 Poor; > 0.2 ≤ 0.4 Fair; >
0.4 ≤ 0.6 Moderate; > 0.6 ≤ 0.8 Good; > 0.8 ≤ 1 Very good. Despite the fact
that this calculation is considered a conservative measure, the inter-coder
agreement was approaching 1, which is perfect.
Affect - κ = 0.949181 Theme - κ = 0.9367972
Agent - κ = 0.9875858 Cause - κ = 0.9510682
For the factors of Agent and Cause, only the distinction between ‘state’
and ‘abstract state of affairs’, and, to a small degree, ‘event’, caused any
disagreement. For the factor of Affect, it was found that ‘emotional pain’
and ‘anxiety’ were on a continuum. These two features differ only in their
degree of emotional engagement. When a continuous feature is treated
categorically, it presents a fuzzy boundary leading to coder disagreement.
The other coder disagreement concerned the distinction between ‘imposi-
tion’ and ‘interruption’. Again, there is a fine line between the two affects.
Although it may be difficult to claim that they form a continuum, there is
obviously similarity between having to do something that one does not
want to do and not being able to do something that one wants to do.
In this section, we have overtly described the subjective features that are
annotated. It was argued in the previous section that this stage in itself per-
mits a kind of verification which is sometimes less evident in purely intro-
spective research. Although good introspective studies clearly identify the
criteria used in analysis, by systematically applying those criteria many
hundreds of times to randomly chosen examples, they are refined. It is the
process of annotation that leads to a clearer analysis. Other than the use of
multiple analysts, there is no inherent reason that this stage of the analysis
would be any more accurate than any traditional introspection-based study.
However, we now have one very important advantage – a fine-grained an-
notated dataset. In this study, the sample consists of 650 examples that can
be examined for tendencies in usage, especially the interaction of different
 Objectivity and verification in usage-based Cognitive Semantics 251

factors such as the effects of formal variation and sociolinguistic variation

in semantic structure.

4. Results

4.1. Exploratory statistics – Correspondence Analysis

Multiple Correspondence Analysis is a dimension reduction technique

3
similar to Multi-dimensional scaling. It is a simple method that offers in-
sights into the interaction of multiple factors, but does not offer any means
for testing the statistical significance of the associations it reveals. Despite
this limitation, it represents a useful tool for identifying the patterns in the
data, patterns that would be difficult or impossible to identify heuristically.
The analysis works on a simple principle; it calculates a table of co-
occurrences of the different features in question and then converts these
figures to relative distances. The plot is interpreted accordingly – proximity
represents high association, distance represents disassociation. This is, of
course, relative and data points may push other points away so that one
feature may be highly associated with another, even though they are not
placed close together because one of the points is being “pushed” away
from a third and unrelated feature. For this reason, the plots can be difficult
to read, but their analogue representation of the tendencies in correlations
between usage-features avoids giving a misleading picture that we are deal-
ing with discreet senses.

4.1.1. Affect

Affect is obviously one of the most important factors since, as an emotion

term, it is essentially the ‘meaning’ of the lexeme; it is also one of the most
subjective factors under investigation. Let us examine how this factor inter-
acts with the formal variation associated with the lexeme. Restricting the
data to just the verbal examples, we can submit the factor of Affect to a
Correspondence Analysis with the Grammatical Construction. Recall that
the five basic verbal constructions include the Direct Transitive Cx,
Oblique Transitive Cx, Oblique Reflexive Transitive Cx, Complementary
Clause Cx, and the Predicative Cx. Figure 1, below, visualises the results of
the analysis.
252 Dylan Glynn

(i) Oblique Transitive Cx – ‘interruption’ and ‘imposition’

Figure 1 gives us clear results where three distinct clusters emerge. In the
first cluster (i), bottom left, the Oblique Transitive Construction is placed as
distinct from the other constructions. This placement is of interest for three
reasons. First, the fact that it is distinctly placed far from the other clusters
means that relative to the factor of Affect, this construction is used in a
specific way. In other words, it is a form–meaning pair. The two Affect
features are located between it and the Direct Transitive Construction,
which means that they are not unique to the Oblique Transitive Construc-
tion, but on occasions also occur with the direct form. However, the relative
proximity shows their high association with the Oblique form.

energy
physical pain Predicative Cx
emotional pain Oblique Reflexive Cx
anxiety (ii) Complement Clause Cx
Direct Transitive Cx

(iii)

( i)

Figure 1. Correspondence Analysis – Construction and Affect

Second, the Oblique Transitive and the Direct Transitive are the two con-
structions that profile the Agentive Frame of BOTHER. Although the Affects
of ‘imposition’ and ‘interruption’ are clearly more associated with the use
of the Oblique form, it is with the other Agentive construction that they
share some association. In other words, although the Oblique and Direct
Transitive Constructions are distinct, they are also semantically similar,
relative to the factor of Affect. The Affect factor divides the plot into left
 Objectivity and verification in usage-based Cognitive Semantics 253

and right halves – the constructions profiling an Agentive argument struc-

ture appear on the left, and the constructions profiling a Predicative struc-
ture on the right. Third, these two semantic features, ‘imposition’ and ‘in-
terruption’, are similar. Indeed, this was one of the two points of inter-coder
disagreement for this factor. Therefore, that these highly similar and diffi-
cult to distinguish semantic features should behave in a similar manner
relative to the formal variables of construction type verifies the accuracy of
the semantic feature analysis. Example (6) is typical of the kind of occur-
rence this association captures.
(6) a. It’s great because now my sister can’t bother me with her little annoy-
ing-as-fuck friends upstairs anymore. (Oblique Trans Cx and ‘interrup-
tion’)
b. It’s not subconiously anymore either and she told me she couldn’t stop.
and people wouldn’t stop asking her and bothering her about it and I kept
telling them to leave her alone that they were making her feel worse.
(Oblique Trans Cx and ‘imposition’)
These examples also show why there was difficulty in distinguishing the
Affect features of ‘imposition’ and ‘interruption’.

(ii) Direct Transitive Cx – ‘physical pain’, ‘emotional pain’, and ‘anxiety’

The second cluster in the top left of the plot is the Direct Transitive associ-
ated with the Affect features of ‘physical pain’, ‘emotional pain’, and ‘an-
xiety’. This cluster is interesting for the same three reasons as the previous.
First, the semantic features render this a form-meaning pair, distinct from
the other constructions. Second, the construction still shares the left side of
the plot with its daughter construction, the Oblique Transitive Construction,
linked by the features of ‘imposition’ and ‘interruption’. Third, the seman-
tic features that are associated with this form are similar and form a coher-
ent meaning cluster. Indeed, just as for ‘interruption’ and ‘imposition’, the
other difficulty for inter-coder agreement was the distinction between ‘an-
xiety’ and ‘emotional pain’. Both constructions cluster with semantic fea-
tures in an intuitively sound way. It takes a lot of context to annotate such
subjective features. Only short excerpts are included, hopefully long en-
ough to represent the emotions at stake.
(7) a. I had such a great time and I missed hanging out with those peeps well
some of them at least. I don’t know but for some it just doesn’t bother me
anymore maybe it’s their fucking attitude. (Direct Transitive Cx and ‘an-
xiety’)
254 Dylan Glynn

b. I am trying not to be too sensitive because I know he doesn’t mean to

hurt me. Normally comments don't bother much I've certainly heard much
worse... I guess it just hurts me when he does it because I love him so
much. (Direct Transitive Cx and ‘emotional pain’)
(iii) Oblique Reflexive, Complement Clause, Predicative Cxs – ‘energy’
The third cluster of the top right groups all the Predicative Frame Construc-
tions with a single Affect feature, ‘energy’. The use of energy was the most
common Affect with 314 out of 628 verbal occurrences and it seems highly
associated with the Predicative Frame Constructions. Finally, it should be
stressed that the results of this first Correspondence Analysis add weight to
the distinction between a Predicative and Agentive Frame of BOTHER. Ex-
ample (5a) above is typical of the Predicative Frame Constructions, exam-
ples (8a) - (8c) correlate the Complement Clause and Oblique Reflexive
Cxs with the Affect feature of ‘energy’.
(8) a. i didn’t bother with breakfast as i was being picked up at 11:30 Com-
plement Clause and ‘energy’)
b. why do I bother to update my journal? (Complement Clause and ‘en-
ergy’)
c. hes sweet keeps askin me stuff about the essay though and i dont actu-
ally have a clue what its about so im not actually much use so dont know
why he bothers lol! (Oblique Reflexive Cx & ‘energy’)

4.1.2. Agent

Let us consider another Correspondence Analysis, this time with a third

factor added, that of Agent. The constructions behave in the same manner,
grouping together relative to the semantic features along the lines of Agen-
tive and Predicative Frames. Thus, in order to render the plot more legible,
the constructional categories are conflated to Predicative and Agentive. The
Correspondecne Anlsyis in Figure 2 reveals a high degree of correlation
between the Agent and the Affect. Most of the correlations are immediately
interpretable and intuitively sound.

(i) Humans, ‘interruption’ and ‘imposition’

The first cluster (i) shows a strong and distinct correlation between ‘inter-
ruption’ and ‘imposition’ with ‘Non-Specified Humans’ and ‘Specified
Humans’. The association there should be self-explanatory. Example (6),
above, serves well to represent this cluster of features.
 Objectivity and verification in usage-based Cognitive Semantics 255

(ii) Agentive Frame Cxs – ‘things’, ‘states’, ‘pain’, and ‘anxiety’

Likewise, in cluster (ii), that ‘emotional pain’ and ‘physical pain’ should be
the Affect resulting from Agents such as ‘things’, ‘states’ and ‘abstract
states of affairs’ is intuitively sound. These two kinds of Agent and Affect
represent two of the central meanings of bother. These constructions and
semantic features were amongst the most common in the sample.
(9) a. my confidence aint exactly full.. but i stil have the courage 2 go on
stage and dance cz i said 2 alison i would. But then tht lot.. had 2 go and
shout stuff while i wis on stage.. i really wanted 2 walk off but cz of
alison i didnt.. they can get 2 fuk.. im not gona let it bother me.. thrs 2
many otha things on my mind wifout tht lot anol!! (Agentive Frame Cx,
‘abstract state of affairs’, & ‘anxiety’)
b. Rob is all stressy coz he has this red rash bloody thing on his face n dnt
want lauren 2 c him :P awwww..he dnt reilise tht she loves him sooo
much it dnt bother her wot is on his face lol :) (Agentive Frame Cx, ‘con-
crete thing’, & ‘anxiety’)

energy
Event
Predicative Bother
physical pain
emotional pain
anxiety
State (ii) (iii)
AbSoA
Aggentive Bother

(i)

Figure 2. Correspondence Analysis – Construction, Affect and Agent

(iii) Predicative Frame Cxs – ‘events’ and ‘energy’

Lastly, we learn that people use the word bother combined with a Predica-
tive Frame Construction when talking about the use of energy that results
from having to do some ‘event’. For example:
256 Dylan Glynn

(10) i might be going to blackpool for the weekend i dont know i dont really
do anything on my birthday because its just nothing if you think about it
because what you get a few gifts and then your just a day older than you
were the day before that so why do they bother with birthdays?
The two Correspondence Analyses presented here are only a sample of a
wider range of analyses that were performed. They have focused on two of
the semantic factors under investigation. There is a much wider range of
features, semantic and otherwise, that are informative here. However, for
practical reasons, we restrict the study to just these dimensions. These two
factors alone, combined with a coarse-grained level of constructional an-
alysis have produced three clusters of form-meaning features that could be
argued to represent lexical senses. Importantly, these senses are associated
with certain syntactic patterns. Results such as these underline the fact that
lexical senses are not discreet, but continuous and multidimensional. Fur-
thermore, they are often drawn towards certain formal variants of a word’s
use and should not be seen as pockets of meaning inside words waiting to
be activated by situation–context. In Cognitive Linguistics, we have dis-
mantled the syntax–lexicon distinction for good reason - the meaning of
words is wrapped up with the semantics of grammatical variation.

4.1.3. Hypothesis

These exploratory analyses have given us a clear and testable hypothesis.

The Correspondence Analyses revealed a semantically motivated distinc-
tion between two sets of syntactic patterns, licensed by the verb bother. We
have termed these patterns Agentive and Predicative Constructions. Con-
firmatory statistics can help here in three ways. Firstly, it can corroborate
the findings of the Correspondence Analyses. Secondly, it can give us a
probability score that our findings are not chance. In other words, if we
were to repeat this analysis another 100 or 1,000 times, it can tell us what
are the chances that we would obtain similar results. Thirdly, it will offer us
information on how explanatorily powerful our analysis is. That is, quanti-
tatively, how often can we predict the usage, or speaker’s choice, of the
construction–verb pairing, given our factors of analysis and their applica-
tion to the data?
 Objectivity and verification in usage-based Cognitive Semantics 257

4.2. Confirmatory statistics – Logistic Regression Analysis

Unlike Correspondence Analysis, regression modelling is a confirmatory

technique. Since confirmatory modelling gives probability scores and also
calculates the explanatory power of the model, it is more complicated than
exploratory techniques such as Correspondence Analysis or Hierarchical
Cluster Analysis. This is because one must test the model for a range of
possible problems, all of which or a combination of which may lead to false
predictions of probability or a misleadingly good (or bad) interpretation of
the accuracy of the linguistic analysis.
The principle of a Logistic Regression Analysis is quite simple. Given a
binary response variable, such as the speaker’s choice of a Predicative
Frame Construction over an Agentive Frame Construction, the model cal-
culates how accurately one may ‘predict’ that choice, with the given lin-
guistic analysis of the data. Accordingly, the model is made up of predictor
factors, each of which offers information to help the model accurately pre-
dict the outcome. In simpler terms, imagine having a list of examples and
your analysis, then hiding whether the example is Agentive or Predicative.
With this hidden, try to guess, based on the feature analysis of the exam-
ples, whether it is Agentive or Predicative. How often you get the answer
correct is the explanatory power of the model. Obviously, if one can accu-
rately predict the speaker’s choice in such a situation, one has analysed the
data accurately and sufficiently.
We saw in the Correspondence Analyses that both the semantic factors
of Affect and Agent were important in understanding the difference be-
tween the Predicative and Agentive Constructions. We cannot use the
Cause factors to distinguish between Predicative and Agentive because all
examples of Predicatives were either ‘specified human’ or ‘non-specified
human’. The one-to-one correlation here is obviously unhelpful in multifac-
torial modelling. This factor is insightful solely for the Agentive Construc-
tions. For reasons of brevity, we will not consider the results of those ana-
lyses. Another Correspondence Analysis, not presented, revealed that the
other highly subjective factor, Theme (Topic of Discourse), also interacts in
informative ways with both the semantics and the constructions. We will
include this factor in the regression modelling.
Several models are run with variation combinations of these semantic
factors as well as other factors, such as the presence of humour, the re-
gional variation and certain formal factors, such as tense, aspect, and mood.
After comparing a range of models, the most significant and explanatory
258 Dylan Glynn

5
ones are selected. Although a combination of Affect and Agent was the
most powerful, multicollinearity was identified between the two factors.
Multicollinearity is a serious problem for Logistic Regression and occurs
when 2 or more of the factors are too highly correlated, that is, they predict
in too similar a way. This exaggerates the accuracy of the model and may
lead to false predictions. Of the two remaining models, Affect and Theme
or Agent and Theme, Agent proved to be a better predictor than Affect. The
analysis was performed in R and the model is presented below:
Binomial Logistic Regression
Construction ~ Theme + Agent

Deviance Residuals:
Min 1Q Median 3Q Max
-1.9986 -0.6330 0.2825 0.6712 2.0376

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.319657 0.384230 0.832 0.405442
Theme_Family -2.535217 0.699139 -3.626 0.000288 ***
Theme_Health -17.879458 522.211504 -0.034 0.972687
Theme_Pers_Entertain -2.059555 0.727649 -2.830 0.004649 **
Theme_Pers_Misc -1.825373 0.380873 -4.793 1.65e-06 ***
Theme_Pub_Entertain -1.984569 0.635572 -3.122 0.001793 **
Theme_Religion -2.261638 0.569255 -3.973 7.10e-05 ***
Theme_Soc_Misc -1.946970 0.551445 -3.531 0.000415 ***
Theme_Uni-School -1.349863 0.638618 -2.114 0.034539 *
Agent_Event 2.881617 0.259185 11.118 < 2e-16 ***
Agent_HumNSp 0.006164 0.705198 0.009 0.993026
Agent_HumSp 1.262592 0.421347 2.997 0.002730 **
Agent_State 0.504517 0.403608 1.250 0.211293
Agent_Thing 1.018456 0.371297 2.743 0.006089 **
---
Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

Null deviance: 866.28 on 627 degrees of freedom

Residual deviance: 586.06 on 614 degrees of freedom
AIC: 614.06
Number of Fisher Scoring iterations: 15
Summary of Model Preditive Power of Model
Model L.R.: 280.22 Pseudo R2: 0.481 (0.468 penalised)
D.f: 13 C: 0.844 (0.843 penalised)
P: 0 Somers' Dxy: 0.687 (0.687 penalised)

Adding interactions between the predictors did not improve the model;
changing the order in which the levels were introduced did not affect the
results in a considerable way, and comparing the AICs did not suggest that
a stepwise regression improved the model. In case of possible over-fitting
in the model, a Penalised Maximum Likelihood measure was determined
and added. This simply adds a penalty factor to the estimations.
 Objectivity and verification in usage-based Cognitive Semantics 259

Before we interpret the results, let us consider the accuracy of the over-
all model. Beneath the table of coefficients and degrees of freedom, we
have a small list of figures. We want to know what proportion of the vari-
ance the model explains and how well it can predict the outcome as either
Agentive or Predicative. The Nagelkerke Pseudo R2 is calculated from log-
likelihood ratios. Any figure above 0.3 is sign of a predictive model. How-
ever, a true R2 calculation is not possible for Logistic Regression and many
scholars consider this an unreliable score save when one is comparing dif-
ferent models. It was one of the key factors in choosing this model over the
Theme and Affect model. The C-score is perhaps the most important here
and is an index of the correlation between predicted probability of expected
response and actual response. For the C-score, 0.5 is pure chance, 1 is per-
fect. Any value over 0.8 is an explanatory result. The score at C-0.844 is a
strong result. The Dxy is another way of measuring C, based on a rank cor-
relation of probabilities and responses. For this score, 0 is randomness and
1 is perfect prediction. Our score of 0.687 also represents a strong result.
The model was checked for issues of multicollinearity. The variance in-
flation factors were calculated and no problems were found. Only the level
‘ThemePers_Misc’ revealed a figure of any consequence (3.056959). This
is, however, still well beneath the most conservative figures for multicol-
linearity. Over-dispersion does not appear to be a serious problem - residual
Deviance and Degrees of Freedom are relatively close, and a Chi-square
confirms the two figures are not significantly different (586.06 on 614, p -
0.7854219 Chi-square).
Having established that the model is statistically significant, explanator-
ily powerful, and will not produce false predictions, we can interpret the
results. The column on the left is a list of the different linguistic features,
belonging to the two factors, Agent and Theme, that go into the model and
predict the outcome of an example as Predicative or Agentive. On the far
right, there is a list of probability scores. These rate statistical significance
and should not be confused with explanatory importance. They should be
read as typical p-values. The stars are there to facilitate quick reference.
Once we have identified which of the levels, or linguistic features, are sig-
nificant, we may look at the estimates of the coefficients. This is the list of
figures on the left, where each coefficient is accompanied by its estimated
standard error and a Wald z-value. In the first list, a negative number pre-
dicts for one outcome and a positive coefficient predicts for the other. In
our case, a positive figure means that it is contributing to the prediction of a
Predicative Construction, and a negative, to an Agentive Construction. For
260 Dylan Glynn

these scores, the greater or smaller the number, the more important this
feature is in predicting the outcome.
Remembering that a positive figure predicts for Predictive Construc-
tions, let us consider the list of coefficients. Agent ‘event’ is clearly the
most important. This confirms the high association seen in the Correspond-
ence Analysis. The Agents ‘Specified Human’ and ‘Thing’ are also import-
ant predictors of the Predicative Construction. For the Theme features, we
see that the coefficients are negative numbers, which means they contribute
to predicting an Agentive Frame Construction outcome. The first level, or
linguistic feature, i.e., Theme ‘Cyber Society’, is not listed because the
other levels are calculated relative to this. All of the listed features are sig-
nificant, save Theme ‘health’. As a rule, any figure higher or lower than +/-
1 is a relatively important predictor. Since all of the significant Themes are
approximately -2, we can suppose that each is a reasonably strong indicator
of the Agentive Construction.
This section has shown how confirmatory techniques, based entirely on
highly subjective annotation, not only produce coherent results but results
that can accurately predict the data. With the use of just these two factors,
we are able to accurately predict whether an example will be Predicative or
Agentive for more than two thirds of the examples. This, in turn, confirms
which semantic features are typical of which constructions.

4.3. Summary

This short study has proposed an, albeit incomplete, semasiological map of
the verb bother. It must be remembered that ‘bother’ is an emotion concept
and so identifying, and indeed predicting, in what situations the term is
used puts subjective analysis to the test. The results above come from an
operationalised, verified, and statistically confirmed treatment of an ex-
tremely subjective object.
Relative to three constructions, and with the caveat that these are ten-
dencies in usage, both semantically and as form–meaning pairings, we can
propose three senses of to bother.
 Objectivity and verification in usage-based Cognitive Semantics 261

Sense 1 – Pain
Affect: The experiencer feels hurt by physical pain through emotional
pain to general psychological stress
Agent: Friends and family and other known individuals, states and ab-
stract states of affairs
Form: This use is associated with the Agentive Frame and Direct Tran-
sitive Construction
Sense 2 – Annoyance
Affect: The experiencer feels imposed upon or prevented from doing
what he or she wishes to do
Agent: People, especially people that the experiencer does not person-
ally know
Form: This use is associated with the Agentive Frame and Oblique
Transitive Construction
Sense 3 – Hassle
Affect: The experiencer feels put out by the need to do something that
involves the use of energy
Agent: Events, basically ‘having to do things’
Form: This use is associated with a cluster of constructions where the
Agent and the Patient are encoded by the same actor, termed, in
this study, Predicative Constructions
Confirmatory analysis, at a more coarse-grain level, proved that the differ-
ence between the Agentive (Pain and Annoyance) uses and the Predicative
use was statistically significant. It showed that the factors of Agent and
Thematic Topic of Discourse were the two crucial in features in distin-
guishing their use, or ‘meaning’.

4.4. Theoretical implications

More generally, these findings have two important implications for polys-
emy research:
– No discreet senses. Similar semantic features group together, suggest-
ing “senses” of the lexeme, but without discreet differences
– No pure semasiology. These sense groupings were associated with
certain formal variants of the same word, creating a small onomasi-
ological field ‘within’ the lexeme
262 Dylan Glynn

Firstly, the different senses identified represent tendencies, not discreet

categories. Lakoff (1987: case study 2) assumes distinct lexical senses and
Langacker (1994) overtly argues that senses, like linguistic forms, are ne-
cessarily discreet. The results here suggest that such positions are erro-
neous. They further supports Geeraerts (1993) who has theoretically shown
that a reified understanding of meaning is merely a result of the Structur-
alist framework.
Secondly, we saw how the tendencies towards different uses, or lexical
senses, were associated with different formal variants of the lexeme. Again,
this should come as no surprise for Cognitivists, who assume that variation
in form is motivated by variation in meaning. However, it is interesting to
note that despite this, most cognitive studies of lexical meaning remain at a
very coarse level of analysis of either the lemma or the word. The results
here support Newman (2008; this vol.) and Glynn (2009, 2010), arguing
that the semasiological – onomasiological distinction is not theoretically
tenable and that more fine-grained formal variation must be included in
semasiological study. This runs contrary to Geeraerts et al. (1994) and
Geeraerts (2006) who argue this distinction is fundamental. The results here
suggest that a continuum of granularity exists between the study of semasi-
ology (uses of a single form) and onomasiology (the choice between
forms). This is because, at some fine-grained level (in constructional, mor-
phological, or prosodic variation), there is always a choice between forms
and this choice in form, to some extent, relates to semantic variation.

5. Conclusion

This study has shown that even highly subjective semantic categories such
as the emotional state experienced by a Patient, the Cause of that Affect,
and the Theme of discourse can be operationalised in manual annotation.
Three methodological steps permit sound and meaningful results from even
highly subjective analysis in corpus linguistics.
- Operationalisation
- Inter-Coder Verification
- Confirmatory Statistics
Although these steps do not offer objectivity, they afford a means for veri-
fication and facilitate the empirical cycle of proposing hypotheses and test-
ing them. Step one allows other research to check and improve upon exist-
 Objectivity and verification in usage-based Cognitive Semantics 263

ing analyses; step two increases accuracy; step three offers statistical sig-
nificance and a measure of explanatory power. Although statistics does not
verify the semantic analysis per se, if the analysis was unduly inaccurate,
the modelling would not produce coherent results. In brief, this final step
gives us a way of testing a hypothesis. Since, in Cognitive Linguistics and
in corpus linguistics, we have no grammatical rules or propositional seman-
tics, we cannot use negative evidence to disprove our proposals about how
language works. This does not mean we should not propose hypotheses. On
the contrary, it makes it all the more important that we are explicit about
such things. Statistical significance and the explanatory power of multivari-
ate models are important means for testing those hypotheses.
The study has not attempted to cover the semantic variation of the verb
entirely and it has only included the bare minimum of formal variation.
Moreover, it has excluded most of the sociolinguistic dimensions that need
to be included in such research. Tummers et al. (2005), Grondelaers et al.
(2007, 2008), Glynn (2009, submitted), and Geeraerts & Speelman (2010)
are examples of how such an approach integrates with sociolinguistic pa-
rameters of language. The method employed in this study directly meets
such research and integrating such extralinguistic variables is straightfor-
ward.
Importantly, the results here further confirm two theoretical positions.
Firstly, it was shown that meaning should not be treated as reified senses,
rather as emergent and multidimensional. This supports positions for-
warded by Geeraerts (1993) and Kilgarriff (1997). Secondly, we saw fur-
ther evidence that the semasiological – onomasiological distinction needs to
be placed on a continuum of granularity. Highly schematic studies can
work with formal distinctions at the level of words, morphemes, and con-
structions, but as a study considers more fine-grained semantic structure, it
needs to include other formal variation. Any variation in person, number,
tense, aspect, syntax, mood, prosody and so forth, represents semantic vari-
ation. One can never be sure the meaning identified for a given lexeme
does not change in a different grammatical construction, tense or voice.
Therefore, just as formal types of structure fall on a continuum from lexis
to syntax, so too must the semasiological – onomasiological division. These
results confirm the position of Newman (2008; this vol.) and Glynn (2009,
2010).
To summarise, the strong tendency for corpus-driven research to be shy
of purely semantic analysis is, in itself, a good tendency. Interest in the
method is largely motivated by a desire to move away from introspective
264 Dylan Glynn

methodology. However, as the method matures and as Cognitive Linguist-

ics adopts it, we should be careful not to think that linguistics will ever be a
purely objective science. Linguistics is a social science and the difficulty of
our object of study is precisely its human element. As Cognitive Linguists,
we believe that language is semantically motivated, and as such, at some
stage, we will have to apply these techniques to semantics. The study has
shown, in a small way, why we should think that this will be possible.
Knowing this, Cognitive Semantics in its study of highly abstract gram-
matical semantics, cognitive models, and metaphors should consider adopt-
ing quantitative corpus-driven methods.

Notes

1. This research was completed with a grant from the University of Leuven. I
would like to thank the entire QLVL team at the University of Leuven for
their help. I would also like to thank B. Holmquist, K. Krawczak, and J. van
de Weijer for their help. Oversights, omissions, and errors are entirely my
own.
2. The corpus is based on the LiveJournal personal diary service. The data were
extracted with their permission and the corpus compiled by D. Speelman at
the QLVL research unit, the University of Leuven.
3. Correspondence Analysis has enjoyed less popularity within the cognitive
community than other exploratory techniques. Szelid & Geeraerts (2008) and
Glynn (2009; 2010) are examples of its application. Glynn (in press) offers a
summary of the techniques and a tutorial on how it is used.
4. Within Cognitive Linguistics, there exists a strong tradition of using Logistic
Regression Analysis. Examples include Heylen (2005), Bresnan et al. (2007),
Grondelaers et al. (2008), Diessel (2008), Speelman & Geeraerts (2010),
Speelman et al. (2010) and, Szmrecsanyi (2010). Speelman (in press) offers
an explanation of its use. Several other studies employing the technique are
included in Glynn & Robinson (in press).
5. See Bayaan (2008), Speelman (in press) for an explanation on the tests for the
validity of a Logistic Regression model.
 Objectivity and verification in usage-based Cognitive Semantics 265

References

Baayen, R. Harald
2008 Analyzing Linguistic Data: A practical introduction to statistics
using R. Cambridge: Cambridge University Press.
Bresnan, Joan, Anna Cueni, Tatiana Nikitina and Harald Baayen
2007 Predicting the Dative alternation. In Cognitive Foundations of Inter-
pretation, Gosse Bouma, Irene Kraemer and Joost Zwarts (eds.), 69–
94. Amsterdam: Royal Netherlands Academy of Science.
Divjak, Dagmar
2006 Delineating and structuring near-synonyms. In Corpora in Cognitive
Linguistics. Corpus-based approaches to syntax and lexis, Stefan
Gries and Anatol Stefanowitsch (eds.), 19-56. Berlin / New York:
Mouton de Gruyter.
Diessel, Holger
2008 Iconicity of sequence. A corpus-based analysis of the positioning of
temporal adverbial clauses in English. Cognitive Linguistics 19:
457–482.
Geeraerts, Dirk
1993 Vagueness’s puzzles, polysemy’s vagaries. Cognitive Linguistics 4:
223-272.
2006 Words and Other Wonders: Papers on lexical and semantic topics.
Berlin / New York: Mouton de Gruyter.
2007 Contextualisation and methodology: 30 years of Cognitive Linguist-
ics, Plenary Lecture, 10th International Cognitive Linguistics Con-
ference, Kraków.
2010 Recontextualizing grammar: Underlying trends in thirty years of
Cognitive Linguistics. In Cognitive Linguistics in Action: From
theory to application and back, Elżbieta Tabakowska, Micha Choin-
ski and Łukasz Wiraszka (eds.), 71–102. Berlin / New York: Mou-
ton de Gruyter.
this vol. The doctor and the semantician.
Geeraerts, Dirk, Stefan Grondelaers and Peter Bakema
1994 The Structure of Lexical Variation. Meaning, naming, and context.
Berlin / New York: Mouton de Gruyter.
Glynn, Dylan
2009 Polysemy, Syntax, and Variation. A usage-based method for Cogni-
tive Semantics. In New Directions in Cognitive Linguistics, Vy-
vyan Evans and Stephanie Pourcel (eds.), 77-106. Amsterdam /
Philadelphia: John Benjamins.
2010 Synonymy, lexical fields, and grammatical constructions. A study in
usage-based Cognitive Semantics. In Cognitive Foundations of Lin-
266 Dylan Glynn

guistic Usage-Patterns, Hans-Jörg Schmid and Susanne Handl

(eds.), 89-118. Berlin / New York: Mouton de Gruyter.
in press Multiple Correspondence Analysis. Exploring correlations in multi-
factorial data. In Polysemy and Synonymy. Corpus methods and ap-
plications in Cognitive Linguistics, Dylan Glynn and Justyna Robin-
son (eds.). Amsterdam / Philadelphia: John Benjamins.
submitted Semantics of sociolinguistic variation. A quantitative study of dia-
lect effects on polysemy. Annual Review of Cognitive Linguistics.
Glynn, Dylan and Justyna Robinson (eds.)
in press Polysemy and Synonymy. Corpus methods and applications in Cog-
nitive Linguistics. Amsterdam / Philadelphia: John Benjamins.
Gries, Stefan and Anatol Stefanowitsch
2004 Extending collostructional analysis: a corpus-based perspective on
‘alternations’. International Journal of Corpus Linguistics 9: 97-
129.
Gries, Stefan
2006 Corpus-based methods and Cognitive Semantics: The many mean-
ings of to run. In Corpora in Cognitive Linguistics. Corpus-based
approaches to syntax and lexis, Stefan Th. Gries and Anatol Ste-
fanowitsch (eds.), 57-99. Berlin / New York: Mouton de Gruyter.
Grondelaers, Stefan and Dirk Speelman
2007 A variationist account of constituent ordering in presentative senten-
ces in Belgian Dutch. Corpus Linguistics and Linguistic Theory 3:
161-194.
Grondelaers, Stefan, Dirk Geeraerts and Dirk Speelman
2007 A case for a cognitive corpus Linguistics. In Methods in Cognitive
Linguistics, Monica Gonzalez-Marquez, Irene Mittelberg, Seana
Coulson and Michael J. Spivey (eds.), 149-169. Amsterdam /
Philadelphia: John Benjamins.
Grondelaers, Stefan, Dirk Speelman and Dirk Geeraerts
2008 National variation in the use of er “there”. Regional and diachronic
constraints on cognitive explanations. In Cognitive Sociolinguistics:
Language variation, cultural models, social systems, Gitte Kristian-
sen and Rene Dirven (eds.), 153-204. Berlin / New York: Mouton de
Gruyter.
Heylen, Kris
2005 A quantitative corpus study of German word order variation. In
Linguistic Evidence: Empirical, theoretical and computational per-
spectives, Stephan Kepser and Marga Reis (eds.), 241-264. Berlin /
New York: Mouton de Gruyter.
 Objectivity and verification in usage-based Cognitive Semantics 267

Hilpert, Martin
2006 Distinctive collexeme analysis and diachrony. Corpus Linguistics
and Linguistic Theory 2: 243-57.
2009 The German mit-predicative construction. Constructions and
Frames 1: 29-55.
Kilgariff, Adam
1997 I don’t believe in word senses. Computers and the Humanities 32:
91-113.
Kövecses, Zoltán
1986 Metaphors of Anger, Pride, and Love. A lexical approach to the
structure of concepts. Amsterdam / Philadelphia: John Benjamins.
Lakoff, George
1987 Women, Fire, and Dangerous Things. What categories reveal about
the mind. London: University of Chicago Press.
Langacker, Ronald
1994 The limits of continuity: Discreetness in Cognitive Semantics. In
Continuity in Linguistic Semantics, Catherine Fuchs and Bénard
Victorri (eds.), 9-21. Amsterdam / Philadelphia: John Benjamins.
Newman, John
2008 Aiming low in linguistics: Low-level generalizations in corpus-
based research. Proceedings of the 11th International Symposium on
Chinese Languages and Linguistics, National Chiao Tung Univer-
sity, Hsinchu, Taiwan. Available from <http://johnnewm.
jimdo.com/downloads.php>, accessed 10 January, 2010.
this vol. Balancing acts: Empirical pursuits in Cognitive Linguistics.
Peirsman, Yves, Kris Heylen and Dirk Geeraerts
2010 Applying word space models to sociolinguistics. In Advances in
Cognitive Sociolinguistics, Dirk Geeraerts, Gitte Kristiansen and
Yves Peirsman (eds.), 111-138. Berlin / New York: Mouton de
Gruyter.
Sandra, Dominiek and Sally Rice
1995 Network analyses of prepositional meaning: Mirroring whose mind -
the linguist’s or the language user’s? Cognitive Linguistics 6: 89-
130.
Schmid, Hans-Jörg
2000 English Abstract Nouns as Conceptual Shells. From corpus to cog-
nition. Berlin / New York: Mouton de Gruyter.
Speelman, Dirk and Dirk Geeraerts
2010 Causes for causatives: The case of Dutch ‘doen’ and ‘laten’. Causal
Categories in Discourse and Cognition, Ted Sanders and Eve
Sweetser (eds.), 173-204. Cambridge: Cambridge University Press.
268 Dylan Glynn

Speelman, Dirk, José Tummers and Dirk Geeraerts

2010 Lexical patterning in a Construction Grammar: The effect of lexical
co-occurrence patterns on the inflectional variation in Dutch attribu-
tive adjectives. Constructions and Frames 1: 87-118.
Speelman, Dirk
in press Modeling near-synonymy with logistic regression. In Polysemy and
Synonymy. Corpus methods and applications in Cognitive Linguist-
ics, Dylan Glynn and Justyna Robinson (eds.). Amsterdam /
Philadelphia: John Benjamins.
Stefanowitsch, Anatol and Stefan Th. Gries
2003 Collostructions: Investigating the interaction between words and
constructions. International Journal of Corpus Linguistics 8: 209-
43.
this vol. Empirical Cognitive Semantics: Some thoughts.
Szelid, Veronika and Dirk Geeraerts
2008 Usage-based dialectology: Emotion concepts in the Southern Csango
dialect. Annual Review of Cognitive Linguistics 6: 23–49.
Szmrecsanyi, Benedikt
2003 BE GOING TO versus WILL/SHALL: Does syntax matter? Journal of
English Linguistics 31: 295-323.
2010 The English genitive alternation in a cognitive sociolinguistics per-
spective. In Advances in Cognitive Sociolinguistics, Dirk Geeraerts,
Gitte Kristiansen and Yves Peirsman (eds.), 139-166. Berlin / New
York: Mouton de Gruyter.
Talmy, Leonard
2007 Introspection as a methodology in linguistics. Plenary Lecture, 10th
International Cognitive Linguistics Conference, July 2007, Kraków.
2008 Foreword. In Methods in Cognitive Linguistics, Monica Gonzalez-
Marquez, Irene Mittelberg, Seana Coulson and Michael J. Spivey
(eds.), xxi – xxi. Amsterdam / Philadelphia: John Benjamins.
Tummers, José, Dirk Speelman and Dirk Geeraerts
2005 Inflectional variation in Belgian and Netherlandic Dutch: A usage-
based account of the adjectival inflection. In Perspectives on Vari-
ation. Sociolinguistic, historical, comparative, Nicole Delbecque,
Johan van der Auwera and Dirk Geeraerts (eds.), 93-110. Berlin /
New York: Mouton de Gruyter.
Wierzbicka, Anna
1995 Adjectives vs. verbs: The iconicity of part-of-speech membership. In
Syntactic Iconicity and Linguistic Freezes. The human dimension,
Marge Landsberg (ed.), 223-246. Berlin / New York: Mouton de
Gruyter.
 Objectivity and verification in usage-based Cognitive Semantics 269

Wulff, Stefanie, Anatol Stefanowitsch and Stefan Th.Gries

2007 Brutal Brits and persuasive Americans: Variety-specific meaning
construction in the into-causative. In Aspects of Meaning Construc-
tion, Günter Radden, Klaus-Michael Köpcke, Thomas Berg and
Peter Siemund (eds.), 265-281. Amsterdam / Philadelphia: John
Benjamins.
Zeschel, Arne
this vol. Exemplars and analogy: Semantic extension in constructional net-
works.
Beyond the dative alternation:
The semantics of the Dutch aan-Dative

Timothy Colleman

Abstract
One of the basic theoretical/methodological guidelines of Construction Grammar is
that each construction should first and foremost be analyzed in its own right (cf.
Goldberg 2002, inter alia). In accordance with this guideline, the present paper
gives centre stage to the constructional semantics of the Dutch prepositional dative
construction with aan, the nearest equivalent of the English to-dative. Existing
accounts of this construction are always framed in the context of the dative alterna-
tion, i.e. they focus on the schematic semantic contrasts between the aan-dative
and its “competitor”, the Dutch Double Object Construction. The present paper
takes a different perspective: the overall aim is to provide a usage-based analysis of
the semantic structure of the aan-Dative Construction, following the multidimen-
sional approach to constructional semantics outlined in Geeraerts (1998). The se-
mantic analysis starts out from the list of significantly attracted verbs revealed by a
Collexeme Analysis (Stefanowitsch and Gries) of the lexical material filling the
construction’s verb slot in a one-million-word corpus of contemporary Dutch
newspaper language.
Keywords: aan-Dative Construction, argument structure, constructional semantics,
multidimensional approach, dative alternation, Collexeme Analysis, Dutch

1
1. Introduction

As illustrated in (1) and (2), Dutch displays the phenomenon which is

known as dative shift or the dative alternation in the linguistic literature
(see, e.g., Green 1974, Givón 1984, Pinker 1989). A variety of verbs of
giving, as well as verbs from a number of other, semantically related
classes, can be used in two different grammatical constructions, resulting in
pairs of semantically similar clauses.
(1) Jan heft zijn broer een boek gegeven / overhandigd / verkocht / beloofd.
John has his brother a book given / handed /sold / promised.
‘John has given / handed / sold/ promised his brother a book.’
272 Timothy Colleman

(2) Jan heft een boek aan zijn broer gegeven / overhandigd / verkocht /
beloofd.
John has a book to (aan) his brother / given / handed / sold / promised.
‘John has given/handed/sold/promised a book to his brother.’
The instances in (1) exemplify the Double Object Construction (henceforth:
DOC), which encodes both the theme and the recipient participant of the
denoted ‘possessional transfer’ event as zero-marked NP objects. The in-
stances in (2) exemplify the so-called prepositional dative construction, in
which the theme participant appears as a zero-marked object but the recipi-
ent participant is encoded as an NP in the complement of the preposition
2
aan (cognate with English on, German an). This alternation pattern is of
course well-known from the grammar of English, which exhibits variation
between a Double Object Construction and a prepositional dative construc-
tion as well, see the English glosses in (1) and (2).
The overall aim of the present paper is to provide a usage-based seman-
tic analysis of the Dutch argument structure construction in (2), which, for
ease of reference, shall be referred to as the aan-Dative. The relation be-
tween the constructions in (1) and (2) has received a large amount of lin-
guistic attention within various theoretical frameworks, a major research
strand being concerned with the elucidation of the schematic semantic con-
trasts between them. From the Cognitive Linguistics point of view, gram-
mar is inherently semantic: the difference in grammatical form between the
DOC and the prepositional dative is associated with a (subtle) difference in
meaning, which motivates the choices of speakers for either of the two
“competing” constructions in real language. More specifically, the con-
structions are typically analyzed as differing in grammatical imagery or
construal, i.e. although they both serve to encode ‘possessional transfer’
events, they do so in distinct ways, highlighting distinct facets of the de-
noted event (see Langacker 1991: 13-14 and Panther 1997, inter alia, for
analyses along these lines of the semantic relation between the English
DOC and the to-dative and see Section 2 for similar analyses of the Dutch
facts). In contrast to and as an elaboration on such approaches, the present
paper gives centre stage to the constructional semantics of the Dutch aan-
Dative, true to the Construction Grammar tenet that each construction
should first and foremost be analyzed in its own right (cf. Goldberg 2002).
That is, rather than focusing immediately on the semantic properties which
distinguish the aan-Dative from the DOC, I will examine the semantic
range of the aan-Dative, through an analysis of real language data from a
corpus of contemporary Dutch newspaper language. Moreover, this inher-
 The semantics of the Dutch aan-Dative 273

ently semantic question will be addressed using quantitative analysis. In

Section 3, the “Collexeme Analysis” method introduced by Stefanowitsch
and Gries (2003) will be used to identify the verbs which are significantly
attracted to the aan-Dative’s verb slot in the corpus data. The results from
this corpus-based test will serve as the starting point for a preliminary in-
vestigation of the construction’s semantic structure in Section 4. It will be
shown that only a subset of the verbs which are revealed by the quantitative
analysis to be significantly attracted to the aan-Dative are verbs of giving.
Following the multidimensional approach to constructional semantics ad-
vocated in Geeraerts (1998), I will argue that the semantics of the construc-
tion is built around the concept of ‘caused contact’ and discuss a number of
extensions along two different dimensions of semantic variation, viz. the
exact nature of the ‘contact’ relation involved and the direction of the de-
noted event. Though the present paper does not develop this multidimen-
sional semantic analysis in a quantitative way, it lays the groundwork for
such an enterprise. In addition, as will be argued in more detail below, it
provides an important first step towards a quantitative multifactorial study
of the dative alternation in Dutch which includes an appropriate set of se-
mantic variables.

2. Existing accounts of the aan-Dative: A preoccupation with

‘distinctive features’

Just like the corresponding English phenomenon, the Dutch dative alterna-
tion has attracted a considerable amount of linguistic attention. In Dutch
linguistics as well, many authors take the alternation between the DOC and
the prepositional dative to be a basically semantic phenomenon, associating
the two constructions with related but not identical meanings, in accordance
with the general principle assumed by most cognitive and functional ap-
proaches to language that in human language, a difference in grammatical
form typically signals a difference in meaning (see e.g. Goldberg’s ‘Princi-
ple of No Synonymy’, 1995: 67). Existing semantic accounts more often
than not rely on intuitions about the meanings of construed minimal pairs
of double object and aan-Dative clauses for their evidence. A well-known
example of this approach is found in Van Belle and Van Langendonck
(1996). In their view, the major semantic contrast between the alternating
constructions resides in the fact that the DOC highlights the involvement of
the recipient participant as an interested party, whereas the prepositional
274 Timothy Colleman

construction does not highlight this involvement but rather highlights the
(spatial) transfer aspects of the scene. Among the several observations they
adduce in support of this general hypothesis is the sentence pair in (3) with
the expression iemand zijn arm aanbieden ‘to offer sb one’s arm’. As this
denotes a situation which clearly involves granny as an interested party but
not as the goal of an actual spatial transfer, the DOC in (3a) is the
unmarked option. The aan-Dative clause (3b), by contrast, “suggests a ra-
ther unpleasant interpretation, viz. a physical transfer of a cut-off arm to a
recipient” (Van Belle and Van Langendonck 1996: 239).
(3) a. Vader bood oma zijn arm aan.
‘Father offered granny his arm.’
b. Vader bood zijn arm aan oma aan.
‘Father offered his arm to granny.’
Van Langendonck (2000: 85) summarizes the main semantic contrast as
follows: “[I]t can be stated that an IO-NP stresses the involvement of the
human roles, whereas a PP tends to emphasize the transfer, motion, or ori-
entation embodied by the Goal role”. Similar semantic claims are to be
found in Janssen (1997), Schermer-Vermeer (1991, 2002) and Duinhoven
(2003), inter alia. While the details of these proposals vary, they all agree
that the DOC and the aan-Dative highlight different aspects of the complex
‘give’ scene, with e.g. the relative degree of affectedness or involvement of
the theme and recipient participants and/or the presence or absence of a
spatial transfer constituting important semantic determinants of the dative
alternation. As it happens, very much the same set of semantic notions
plays an important role in many studies of the English alternation (see Col-
leman 2009a for a corpus-based study of the dative alternation in Dutch and
Colleman and De Clerck 2009 for further discussion of existing semantic
hypotheses about the alternation in English and Dutch).
It is by no means my intention to present this particular approach to the
dative alternation as completely misguided, nor to minimize its results. On
the contrary, there is no lack of interesting semantic hypotheses in the
above-mentioned studies, which will have to be taken into account by any
study which aspires to provide an exhaustive overview of the various pa-
rameters driving the dative alternation. Still, it is obvious that as a result of
this methodological focus on minimal pairs and alternating verbs, the pic-
ture of the aan-Dative’s semantics painted in the above studies is necessa-
rily incomplete. While we may have some ideas about the semantic proper-
ties which distinguish the aan-Dative from the DOC, such a list of
 The semantics of the Dutch aan-Dative 275

“distinctive features” should not be mistaken for a full account of the aan-
Dative’s constructional semantics.
One of the basic theoretical/methodological tenets of Construction
Grammar, most vigorously advocated in Goldberg (2002), is that each con-
struction should first and foremost be analyzed in its own right. Goldberg
herself lives up to this principle in, for instance, her seminal analysis of the
polysemous semantics of the English DOC, which arranges the various
subsenses of this construction in a radial network built around a prototypi-
cal ‘Agent successfully causes recipient to receive patient’ sense (Goldberg
1995, 2002). Similarly, Geeraerts (1998) and Colleman (2009b) investigate
the semantic range of the Dutch DOC, without any reference to its para-
phrase possibilities. So far, no attempts have been made at charting the
semantic territory covered by the aan-Dative Construction in a similar fash-
ion, i.e. at providing an overview of the types of situations this construction
can be used to encode (regardless of whether these can also be encoded by
the DOC). This is exactly what the present study sets out to do.

3. A Collexeme Analysis of the aan-Dative’s verb slot

3.1. Introduction

In accordance with common practice in the study of argument structure

constructions, in Construction Grammar as well as in other theoretical
frameworks, the constructional semantics of the aan-Dative will be tackled
through a detailed analysis of the lexical fillers of the construction’s verb
slot. A shared characteristic of some of the most well-known proposals
about the semantics of the English DOC, for instance, is that each includes
some kind of inventory of verb classes associated with the construction, i.e.
a categorization of the verbs which are regularly used with double object
syntax into a number of semantic groups (cf. Green 1974, Wierzbicka
1986, Pinker 1989, Goldberg 1995, inter alia). The exact theoretical status
accorded to these verb classes differs, of course, but there seems to be a
communis opinio that they are an important part of speakers’ knowledge of
the DOC. With regard to the construction under investigation here, this
means that we should be able to identify which (types of) verbs, apart from
those exemplified in (2) above – i.e. geven ‘give’, overhandigen ‘hand’,
verkopen ‘sell’, etc. –, are regularly used to fill the aan-Dative’s verb slot
as well. This question will be addressed through a Collexeme Analysis (cf.
276 Timothy Colleman

Stefanowitsch and Gries 2003) of the construction verb’s slot in a one-

million-word sample from a Dutch newspaper corpus (see Section 3.3 be-
low for further elaboration).

3.2. Towards a formal definition of the aan-Dative Construction

Before we proceed to a discussion of the corpus-based analysis and its re-

sults, it should be pointed out that an important prerequisite for any quanti-
tative investigation is a formal definition of the construction to be investi-
gated. That is, the researcher must have a set of formal criteria which
enable him/her to decide in a consistent way which of the real language
examples encountered in the corpus are actual occurrences of the construc-
tion under investigation (and, crucially, which are not). In the case of the
aan-Dative, this is by no means a trivial matter, as it depends on the syntac-
tic status of the aan-phrase in clauses such as (2) above, which is a subject
of debate.
In order to give an idea of the issues at stake, it is useful to start with a
brief discussion of the corresponding English construction with to. A wide-
spread semantic hypothesis about the English dative alternation is that the
DOC is somehow associated with a ‘caused possession’ meaning while its
to-paraphrase is associated with a ‘caused motion’ meaning. That is, in
(4a), Mary is presented as the person who comes to possess an apple at
John’s instigation, whereas in (4b), she is presented as the goal at the end of
the spatiotemporal path traversed by an apple. This ‘caused motion’ mean-
ing of (4b) is attributed to the semantic import of the allative spatial prepo-
sition to (see Pinker 1989, Langacker 1991, Goldberg 1995, Harley 2002,
inter alia, for different theoretical formulations of this general hypothesis).
(4) a. John gave Mary an apple.
b. John gave an apple to Mary.
This hypothesis implies that the construction in (4b) consists of a verb, an
agent which is linked to subject function, a theme which is linked to direct
object function and an oblique goal argument. Goldberg (1995: 89), for
instance, analyzes (4b) in exactly these terms: it is an example of the so-
called caused-motion construction, which encodes all kinds of situations
where an entity causes another entity to undergo a change of location. In
other words, in this view, (4b) represents exactly the same grammatical
construction as the examples in (5).
 The semantics of the Dutch aan-Dative 277

(5) a. The teacher sent Jimmy to the headmaster.

b. Bob drove the car to London.
Several other accounts of the dative alternation, however, present a nar-
rower view of the to-dative. While I cannot provide an exhaustive discus-
sion of such views here, it is useful to briefly compare the above ‘caused
motion’ approach with the analysis outlined in Davidse (1996). Davidse
offers a semantic account of the dative alternation in terms of independent
contrasts between (i) a ‘donatory’ vs. a ‘destinatory’ perspective and (ii) a
‘benefiter’ vs. a ‘transfer’ model. What matters most in this context, how-
ever, is that she posits a grammatical distinction between the clauses in (6)
and (7) below.
(6) a. I had to give the post to Joaquin. (Davidse 1996: 309)
b. Tim bites the tomato, passes it to Bella. (1996: 324)
(7) a. The brown in Wal-Mart’s logo conveys “a warmer invitation” to shop-
pers. (1996: 315)
b. The couple may pass one of their genetic defects to a child. (1996: 324)
In Davidse’s view, even the clauses in (7) do not represent the same gram-
matical construction as to-clauses with give etc. The formal argument for
considering them as separate constructions is that the clauses in (6) and (7)
have distinct sets of formal alternatives: more specifically, in (6) but not in
(7), this set of alternatives includes the DOC. This is reminiscent of the
traditional analysis of the to-phrases in clauses like (4b) and (6ab) as indi-
rect objects, based on their paraphrasability by a zero-marked NP (see e.g.
Herriman 1995 for a defence of this traditional approach). All of this is just
to show that the English to-dative has been subject to both lumping (Pinker,
Goldberg, etc.) and splitting (Davidse, Herriman, etc.) approaches.
The Dutch aan-Dative poses similar demarcation problems. Depending
on the degree of lumping or splitting, the (a)-clauses in (8) to (10) either
instantiate one and the same argument structure construction, or two or
even three distinct constructions.
(8) a. Hij gaf een cadeautje aan de kinderen.
‘He gave a present to the children.’
*Waar gaf hij een cadeautje? Aan de kinderen.
‘Where did he give a present? To the children.’
c. Hij gaf de kinderen een cadeautje.
‘He gave the children a present.’
278 Timothy Colleman

(9) a. Hij besteedde het geld aan zijn postzegelverzameling.

‘He spent the money on his collection of stamps.’
*Waar besteedde hij het geld? Aan zijn postzegelverzameling.
‘Where did he spend the money? On his collection of stamps.’
c. *Hij besteedde zijn postzegelverzameling het geld.
‘He spent his collection of stamps the money.’
(10) a. Hij hing zijn jas aan de kapstok.
‘He hung his coat on the hatrack.’
Waar hing hij zijn jas? Aan de kapstok.
‘Where did he hang his coat? On the hatrack.’
c. Hij hing zijn jas over een stoel / in de kast / ….
‘He hung his coat over a chair/in the closet/…’
Just like English to, Dutch aan is a spatial preposition (cf. Section 4.3 be-
low). According to Janssen (1976, 1997), the aan-phrase in clauses with
verbs of giving is a locative adjunct. In this view, which is relevantly simi-
lar to the ‘caused motion’ approach to the English to-dative outlined above,
there is no grammatical distinction whatsoever between the above aan-
examples with geven ‘give’, besteden ‘spend’ and hangen ‘hang’: in all
three cases, the verb is combined with a subject, a direct object and a loca-
tive phrase which encodes the new location of the direct object referent.
Van Belle and Van Langendonck (1996), among others, have argued
against this view, advocating the traditional analysis of the aan-phrase in
(8a) as an object. The present paper adheres to this more traditional view as
well. A crucial difference between (8a) and (9a) on the one hand, and (10a)
on the other, is that in the latter case, the aan-phrase can freely be replaced
by phrases with other spatial prepositions or by adverbs such as daar
‘there’ or waar ‘where’. As illustrated in (10b) and (10c), verbs such as
hangen ‘hang’ only select a locative complement, without specifying the
exact form of this complement. This is not the case in the former two
clauses: both with geven ‘give’ and with besteden ‘spend’, aan is a “speci-
fied preposition” in the sense of Huddleston and Pullum (2002: 272ff), i.e.,
it cannot be substituted for by any other preposition, which is indicative of
a special lexical relation between the verb and the preposition (nor can the
aan-phrase be questioned by means of a locative adverb, as shown in 8 and
9b). This property, in our view, justifies the analysis of the aan-phrases in
3
(8) and (9) as objects.
Van Belle and Van Langendonck (1996) and others advocate a further
distinction between indirect objects with aan and prepositional objects with
 The semantics of the Dutch aan-Dative 279

aan (or, at least, they argue that prepositional indirect objects constitute a
specific subclass within the larger category of prepositional objects). Ac-
cording to this distinction, which has been handed down from traditional
grammar as well, the aan-objects in (8a) and (9a) belong to distinct gram-
matical categories because a DOC paraphrase is possible in the former but
not in the latter case, as illustrated in the (c)-clauses above. I will not, how-
ever, follow Van Belle and Van Langendonck in this respect. Obviously,
the use of DOC paraphrasability as a definition criterion is in conflict with
the aim of the present study to investigate the semantic territory covered by
the aan-construction without reference to its double object alternant, since
this would amount to aprioristically restricting the aan-Dative’s semantic
range to a subset of the semantic range of the DOC. This is why the present
analysis lumps together aan-phrases which are traditionally labelled as
indirect object and aan-phrases which are traditionally labelled as preposi-
4
tional object. To sum up, the aan-Dative is formally defined here as fol-
lows: [Sbj [V Obj1 aan-Obj2]], i.e. the verb is combined with a subject, a
direct object and a second object which is introduced by the prepositional
aan. This definition includes the patterns in (8a) and (9a), but excludes the
5
pattern in (10a).

3.3. Collexeme Analysis

Collexeme Analysis is one of the family of so-called collostructional meth-

ods developed in a series of papers by Stefan Gries and Anatol Ste-
fanowitsch (most notably Stefanowitsch and Gries 2003, Gries and Ste-
fanowitsch 2004a, 2004b), all of which are aimed at determining – on the
basis of corpus data – the degree of association between grammatical con-
structions and the lexical items filling their constructional slots, or between
lexical items occurring in various slots of the same construction. I will only
present a brief sketch of the method here, referring to Stefanowitsch and
Gries (2003) for a more detailed description. Collexeme Analysis – which
is sometimes referred to as simple Collexeme Analysis in order to distin-
guish it from other collostructional methods – computes the strength of the
attraction between a given construction and all the lexical items occurring
in a particular slot of that construction in a corpus, on the basis of the co-
occurrence frequencies schematically represented in Table 1 (for a verb X
and an argument structure construction C).
280 Timothy Colleman

If these four frequencies are entered in a two-by-two table and submit-

ted to the Fisher-Yates exact test, the resulting p-value can be interpreted as
a measure of the collostruction strength between X and C, i.e. the degree of
association between them. If this procedure is repeated for all other verbs
attested in C’s verb slot in the corpus, these verbs can be ranked according
to their collostruction strength. The eventual output of a Collexeme Analy-
sis is a ranked list of collexemes, i.e. lexical items which are significantly
attracted to the investigated constructional slot.
Table 5. Input data for the Collexeme Analysis of a verb X and an argument struc-
ture construction C (after Gries, Hampe, and Schönefeld 2005:644)
construction C ¬ construction C Row totals
verb X a b a+b
(= overall lemma freq of X)

other c d c+d
verbs
than X
column a+c b+d (a+b)+(c+d)
totals (=overall freq of C) (total number of argument
structure constructions in the
corpus)

The top collexemes are the lexical items which most typically realize the
investigated constructional slot in actual language use. As usage-based
approaches to language assume that the semantics speakers associate with
abstract grammatical patterns is the result of generalizing over (repeatedly)
encountered instances (see Langacker 2000, Goldberg, Casenhiser, and
Sethuraman 2004, inter alia), it follows that the verbs most closely associ-
ated with a particular argument structure construction should provide a
good indication of the construction’s meaning. For instance, one of the case
studies in Stefanowitsch and Gries (2003) concerns the English DOC, the
most strongly attracted verb of which is give. Interestingly, however, the
remaining verbs in the top 20 or so of strongest collexemes represent sev-
eral different verb classes and in fact cover the full spectrum of construc-
tional subsenses distinguished in Goldberg’s (1995, 2002) radial network
analysis, thus testifying to the construction’s high degree of polysemy. In
the same way, the ranking of significantly attracted lexical items resulting
from a Collexeme Analysis can be used to get a grip on the semantics of
constructions the semantic structure of which has not been examined be-
 The semantics of the Dutch aan-Dative 281

fore. Hilpert (this vol.), for instance, discusses the results from a Collexeme
Analysis of the English gerund complement clause construction, grouping
the gerund-infinitive combinations revealed by the quantitative analysis to
be significantly attracted to the construction (such as learning to read, at-
tempting to murder, intending to cause) into four categories, each of which
represents a particular force-dynamic scenario in terms of Talmy (2000),
i.e. a particular constellation of an agonist and an antagonist with their in-
herent force tendencies. In this manner, the Collexeme Analysis provides
empirical support for an analysis of the constructional semantics which is
firmly rooted in the category of force dynamics. Other constructions which
have been investigated from this collostructional perspective include the
English into-causative (Stefanowitsch and Gries 2003, Gries and Ste-
fanowitsch 2004b), the English as-predicative construction (Gries, Hampe,
and Schönefeld 2005), the Swedish ska and komma att constructions (Hil-
pert 2008), the Dutch DOC (Colleman 2009b), and so on. The present pa-
per sets out to do the same for the Dutch aan-Dative.
It should be added that issues of constructional semantics have also been
tackled through Distinctive Collexeme Analysis, an extension of the
method which is specifically tailored for the investigation of grammatical
alternations (Gries and Stefanowitsch 2004a). Distinctive Collexeme An-
alysis computes the degree of association between lexical items and two or
more functionally similar constructions on the basis of their co-occurrence
and overall frequencies. The output is an ordered list of distinctive collex-
emes for each of the investigated constructions, i.e. of those items which
significantly prefer that particular construction over the other(s). It has been
applied to various grammatical alternations, including the variation be-
tween the go-V and go-and-V constructions in English (Wulff 2006) and
the variation between the futurate present and the werden construction in
German (Hilpert 2008), to give but two examples. Such studies provide
relevant insights about the semantic properties of grammatical construc-
tions as well. This is not the path chosen in the present study, however,
which – for reasons discussed in Sections 1 and 2 above – examines the
semantic range of the aan-Dative in its own right (but see Colleman 2009a
for an application of Distinctive Collexeme Analysis to the dative alterna-
tion in Dutch).
282 Timothy Colleman

3.4. The design of the Dutch database and the results of the test

The method of Collexeme Analysis crucially requires that all instances of

the construction under investigation be extracted from the corpus sample.
There is no readily available syntactically annotated Dutch corpus which
would allow for the automatic retrieval of all clauses with a subject, a direct
object and a prepositional object with aan. However, since all instances of
the construction under investigation by definition contain the preposition
aan, it is possible to build an exhaustive inventory of the aan-Dative
clauses attested in a given corpus on the basis of the results of an automatic
search for all instances of the word form aan. I extracted all instances of
aan from a one-million-word sample from the newspaper component of the
ConDiv corpus of written Dutch (Grondelaers et al. 2000), which was quer-
ied by means of the corpus tool Abundantia Verborum (Speelman 1997).
This corpus sample contained 6,983 instances of aan, including some cases
where aan is part of one of the pronominal adverbs daaraan, hieraan,
eraan or waaraan (= the Dutch equivalents of English forms such as
thereto and whereto). These instances were manually skimmed to identify
all clauses with the relevant syntactic pattern [Sbj [V Obj1 aan-Obj2]].
Needless to say, apart from many unproblematic examples, the original
database of 6,983 clauses with aan also contained several occurrences
where it had to be decided on an item-by-item basis whether or not the
example in question instantiated the above syntactic pattern. In most of
these “problematic” cases, the question boiled down to whether the aan-
phrase could be considered a prepositional object along the lines discussed
in Section 3.1 above. As a rule of thumb, I excluded instances in which the
aan-phrase could easily be replaced by locative phrases with all kinds of
other spatial prepositions or adverbs. This excluded instances with verbs
such as zetten ‘put, set, place’, leggen ‘place, lay’, hangen ‘hang’ and plak-
ken ‘stick to, glue on’, see (11a) for a corpus example (with the wider loca-
tive possibilities of zetten illustrated in the construed examples in 11b and
6
11c).
(11) a. Hoe zou België er uitzien als iedereen wrakken aan zijn voordeur
zette? [GvA]
‘What would Belgium look like if everybody put [car] wrecks on their
front door (i.e., outside their front door)?’
Waar zette hij de autowrakken? Aan zijn voordeur.
‘Where did he put the car wrecks? Outside his front door.’
 The semantics of the Dutch aan-Dative 283

c. Hij zette een autowrak in zijn voortuin/op het dak van het winkelcen-
trum/voor de toegangspoort/….
‘He put a car wreck in his front yard / on the roof of the shopping mall /
in front of the entrance gate / … ’
The pattern in (11a) is related to the pattern in the examples (12) below,
which were excluded from the database because the aan-phrase is not an
object but a resultative phrase: rather than introducing a third participant, it
predicates something of the direct object referent. Other verbs instantiating
7
this resultative pattern include krijgen ‘get’ and helpen ‘help’.
(12) a. De eierdooiers wit kloppen met de suiker en de melk aan de kook
brengen. [GvA]
‘Whip the egg yolks and sugar until light and bring the milk to the boil’
b. Op een aantal plaatsen in Italië was het bovendien gebruikelijk dat het
kersverse paar een vaas aan diggelen sloeg. [GvA]
‘In some places in Italy, it was customary for the newly weds to smash a
vase to pieces.’
After the exclusion of these and other cases in which the aan-phrase is not
part of the relevant syntactic pattern [Sbj [V Obj1 aan-Obj2]], the remaining
database contained 1,612 genuine occurrences of the aan-Dative, featuring
187 different verbs. The overall frequencies of these 187 verbs in the cor-
pus sample were computed by means of combined queries for all their
forms. Finally, by POS-tagging the corpus sample with the MBT POS-
tagger for Dutch (Daelemans, Zavrel, and Berck 1996, Zavrel and Daele-
mans 1999) and counting the verb tags, I arrived at an approximation of the
total number of argument structure constructions in the corpus. All of these
frequencies were submitted to version 3.0 of Stefan Gries’ R-script for
computing collostruction strength (Gries 2004).
Of the 187 verbs entered in the test, 92 verbs turned out to be signifi-
cantly attracted to the construction’s verb slot at the 0.05 level of statistical
confidence. Table 2 presents the thirty verbs identified by the analysis as
the strongest collexemes of the Dutch aan-Dative. English glosses are
added to give an indication of the meanings of these verbs, some of which
will be discussed in more detail in Section 4. As advocated by Gries and
Stefanowitsch in later work, the measure of collostruction strength used in
Table 2 is the negative logarithm to the base of ten of the p-value resulting
from the Fisher-Yates Exact test (see e.g. Gries, Hampe, and Schönefeld
2005: 648).
284 Timothy Colleman

Table 2. Collexemes most strongly attracted to the Dutch aan-Dative

Verb (N of aan-Dative Collostruction Verb (N of aan-Dative Collostruction
occurrences) strength occurrences) strength
1. toevoegen (92) 160,75 16. schenken (20) 27,65
‘add’ ‘give (as a present)’
2. besteden (75) 129,72 17. leveren (29) 25,73
‘spend, devote’ ‘furnish, deliver’
3. geven (122) 78,50 18. binden (17) 24,75
‘give’ ‘bind, tie, unite’
4. koppelen (33) 60,62 19. verkopen (33) 23,72
‘pair, connect’ ‘sell’
5. danken (37) 59,7 20. overmaken (17) 23,19
‘owe, (have to) thank’ ‘transfer, remit’
6. wijten (31) 56,94 21. hechten (15) 22,82
‘blame, ascribe’ ‘fasten, attach’
7. voorleggen (39) 50,07 22. aanpassen (22) 19,84
‘present, submit’ ‘adapt’
8. wijden (26) 45,17 23. afstaan (14) 18,46
‘dedicate, devote’ ‘give up, hand over’
9. overlaten (27) ‘leave to’ 42,89 24. bieden (29) ‘offer’ 18,44
10. vragen (72) ‘ask’ 50 25. verlenen (17) ‘grant’ 18,4
11. toeschrijven (18) 33,03 26. doen (66) 14,35
‘attribute’ ‘do’
12. onderwerpen (20) 31,38 27. toekennen (11) 13,61
‘subject’ ‘award, attribute’
13. overhandigen (21) 30,5 28. toevertrouwen (8) 13,03
‘hand’ ‘entrust, commit’
14. overhouden (20) 29,29 29. ontlenen (7) 11,94
‘keep, be left with’ ‘borrow, derive’
15. overdragen (18) 27,94 30. verbinden (10) 11,91
‘pass on, hand over’ ‘connect, join, link’

Note that the ranking of significant collexemes in Table 2 does not coincide
completely with a ranking in terms of raw frequency of occurrence in the
aan-Dative in the corpus (the raw co-occurrence frequencies are included
in brackets in the table). Collostruction strength presents a more refined
measure of attraction, as, in addition to the raw frequency in which the
investigated verb co-occurs with the investigated construction, it also takes
into account the verb’s use in other constructions as well as the construc-
tion’s use with other verbs. Gries, Hampe and Schönefeld (2005) provide a
thorough discussion of the primacy of collostruction strength over raw fre-
quency of occurrence as a measure of the degree of association between
 The semantics of the Dutch aan-Dative 285

verbs and constructions. With regard to the English as-predicative con-

struction, for instance, i.e., the pattern Sbj + Verb + NP-Obj + as + com-
plement, they show that see is the top verb in terms of raw frequency, while
regard is the top verb in terms of collostruction strength. Obviously, regard
provides a stronger clue for the description of the constructional meaning
than the highly polysemous verb see, which frequently occurs in all kinds
of other argument structure constructions as well. They also report on a
sentence completion experiment in which collostruction strength was
shown to significantly outperform frequency as a predictor of the test sub-
jects’ production preferences. To return to the Dutch aan-Dative, the verbs
koppelen ‘pair, connect’, verkopen ‘sell’ and maken ‘make’ all occur 33
times in the construction in the investigated one-million-word corpus; they
are on a shared 9th place in terms of raw frequency of occurrence. However,
while the former two verbs are revealed to be significantly attracted to the
construction according to the Collexeme Analysis (they are ranked 4th and
19th in Table 2, respectively), the general creation verb maken is not sig-
nificantly attracted to the aan-Dative, i.e. its attested frequency in the aan-
Dative does not differ significantly from the co-occurrence frequency that
might be expected by pure chance alone on the basis of the overall frequen-
cies of maken and the aan-Dative in the corpus. I will base my semantic
analysis of the aan-Dative on verbs of the former kind, which do co-occur
much more often with the construction than expected by pure chance and
which as such provide a strong clue towards its constructional meaning.
As a first general observation in this regard, the ranking in Table 2 il-
lustrates that the aan-Dative has a quite broad semantic range. Several of
the top collexemes are verbs of possessional transfer, which corroborates
that this ‘possessional transfer’ use occupies a prominent position in the
construction’s semantic structure. Examples include the basic verb of giv-
ing geven ‘give’, which is the third collexeme and in terms of raw fre-
quency by far the most frequent aan-Dative verb in the corpus sample, as
well as overhandigen ‘hand’, schenken ‘give (as a present)’, verkopen
‘sell’, verlenen ‘grant’, etc. However, this ‘possessional transfer’ use is by
no means the only prominent use of the aan-Dative: among the top collex-
emes are all kinds of other verbs as well, such as toevoegen ‘add’, besteden
‘spend, devote’, koppelen ‘pair, connect’, wijten ‘blame on, ascribe’, on-
derwerpen ‘subject’, hechten ‘fasten, attach’, etc. These are all non-dative
alternating verbs, which, as a result, are hardly ever taken into account in
existing studies. At first sight, they make up a rather heterogeneous – if not
to say ramshackle – collection of verbs.
286 Timothy Colleman

Closer scrutiny, however, reveals that they cluster into a number of dis-
tinct semantic classes. Needless to say, though Collexeme Analysis pro-
vides information about statistical co-occurrence patterns that is impossible
to come by through introspection or through the investigation of raw corpus
frequencies, the detection of semantically coherent clusters of collexemes
in these data is a matter of human interpretation.
Before I proceed to a discussion of some of the verb classes that can be
discerned in Table 2 and of the associated constructional subsenses or
usages, the next section briefly presents the multidimensional approach to
constructional semantics followed here.

4. The semantics of the aan-Dative: Towards a multidimensional

analysis

4.1. A multidimensional approach to constructional semantics

Geeraerts’s (1998) analysis of the Dutch DOC provides a suitable model

for the study of constructional semantics. Crucially, Geeraerts advocates a
multidimensional approach: the semantic structure of the DOC is analyzed
as consisting of a semantic core ‘beneficial transfer of a concrete entity
from an agent to a recipient’ plus extensions along various dimensions.
Each of these dimensions corresponds to a particular component of the
semantic core. For instance, instead of a transfer of a concrete entity, the
construction can also be used to encode various kinds of metaphorical
transfers, e.g. the transfer of a message in clauses with ditransitive verbs of
communication such as vertellen ‘tell’. Or, for another example, in addition
to beneficial transfers from an agent to a recipient, the construction can also
encode situations where the agent deprives the indirect object referent of a
possession. Building on the analysis outlined in Geeraerts (1998), Colle-
man’s corpus-based (2006) study distinguishes six major dimensions of
semantic variation for the Dutch DOC (also see Colleman 2009b). Import-
antly, individual instances of the construction may combine semantic shifts
along several of these dimensions. The corpus example with ontnemen
‘take away’ in (13) below is a case in point: it combines a shift along the
direction dimension (the indirect object referent, Koss, is the source instead
of the endpoint of the transfer) with a shift from a material to an abstract
transfer (i.e. a world record is not a concrete transferred entity). In this way,
 The semantics of the Dutch aan-Dative 287

(13) differs from the semantic prototype with regard to several semantic
dimensions at the same time.
(13) [Begin december] ontnam hij Koss in Heerenveen diens wereldrecord op
de vijf kilometer. [NRC]
‘Early december in Heerenveen, he took the world record for the 5 kilo-
metres away from Koss.’ (lit. away-took Koss the world record)

4.2. The hypothesized semantic core of the aan-Dative

The first step to be taken towards a similar multidimensional analysis of the

semantic structure of the aan-dative consists in a formulation of the con-
struction’s semantic core. While in the case of the DOC, it is intuitively
clear that the category is built around the concept of ‘caused possession’, it
is less straightforward to pinpoint a similar core concept for the aan-Dative.
However, if we assume that the various uses of prepositions are not un-
related, as is a standard assumption in cognitive approaches to language, it
follows that there must be some connection between the meaning of the
aan-Dative and the basic spatial meaning of the preposition aan (cf. the
many analyses of the English to-Dative which attribute a crucial role to the
allative semantic import of to). However, unlike English to, Dutch aan is
basically locative rather than allative in meaning. Previous studies of the
semantics of Dutch spatial prepositions agree that in its basic sense, aan
denotes a relation of spatial contact or contiguity between two stationary
entities (see, e.g., Cuyckens 1991, Schermer-Vermeer 2002, Beliën 2002).
(14) lists a number of examples from Cuyckens (1991: 223-26) which il-
lustrate this basic spatial sense.
(14) a. De poster hangt aan de muur.
‘The poster hangs on the wall’
b. Hij hield zijn vinger aan de trekker.
‘He held his finger on the trigger’
c. de jas aan de kapstok, het bloed aan zijn handen, de ijspegels aan het dak
‘the coat on the hanger, the blood on his hands, the icicles on the roof’
Syntactically, the aan-Dative is a construction which adds a prepositional
phrase to the monotransitive constellation of a subject plus a single nominal
object. In Goldberg’s version of Construction Grammar, arguments that are
encoded as subject or nominal object are considered to be constructionally
profiled: these are the core arguments which are presented as centrally in-
288 Timothy Colleman

volved in the profiled event (cf. Goldberg 1995: 49). In these terms, the
aan-Dative has two constructionally profiled arguments: it denotes a situa-
tion in which the referent of the NP in subject position acts on the referent
of the NP in direct object position. The aan-phrase adds a third construc-
tional argument, which has a certain role to play in the denoted event as
well, but is less crucially involved than the agent and patient arguments. If
we take the role of this added argument to be related to aan’s basic spatial
semantics, we arrive at the following hypothesized constructional meaning
for the aan-Dative: ‘X does something with Y, resulting in contact between
Y and Z’. That is, the construction involves an agent subject, a patient di-
rect object and an aan-object encoding what could be termed a “contactee”,
with the manipulation of the patient by the agent resulting in a relation of
contact between the patient and the contactee.
As a next step, we need to link this formulation in terms of ‘caused con-
tact’ to the various observed uses of the aan-Dative. Is it possible to posit
plausible semantic links between these various uses and the hypothesized
semantic core? Which dimensions of semantic variation should be distin-
guished in such an analysis? Ultimately, the proof of the pudding has to be
in the eating: the more observed uses that can be linked to this ‘caused con-
tact’ scenario via plausible semantic extensions, the more attractive the
resulting overall analysis built around that concept will be. In the next two
subsections, it will be argued that such an analysis should distinguish at
least two dimensions of semantic variation, viz. the ‘contact’ and ‘direction’
dimensions. It should be stressed at this point that the picture of the aan-
Dative’s semantic structure presented below is multidimensional in an ana-
lytic sense of the term, but that it is not a quantitative multifactorial study.
In much the same way, Geeraerts (1998) presents a multidimensional se-
mantic analysis of a grammatical construction but does not address quanti-
tative aspects of semantic structure. Subsection 4.5 presents a number of
ideas about how the present analysis could be developed in a more quanti-
tative way in future work.

4.3. Semantic variation with regard to the ‘contact’ dimension

It is well known that both in English and in Dutch, the DOC can encode all
kinds of situations in which the agent’s act does not cause the establishment
of a prototypical possession relation between a human possessor and a ma-
terial entity, but rather results in a relation between the object referents that
 The semantics of the Dutch aan-Dative 289

can be understood as some kind of metaphorical possession (see Goldberg

1995 and Geeraerts 1998, inter alia, for examples and discussion). In simi-
lar fashion, a brief glance at the range of verbs in Table 2 suffices to show
that the exact nature of the ‘contact’ relation established by the subject’s
action varies widely among the verbs which most typically realize the aan-
Dative.
First, Table 2 includes a number of verbs of attachment, viz. koppelen
‘pair, connect’, binden ‘bind, tie, unite’, hechten ‘fasten, attach’ and
verbinden ‘connect, join, link’. These verbs can be used to encode the cau-
sation of actual spatial contact between two entities, as illustrated for
verbinden ‘connect, join, link’ in (15a), though it should be added that the
majority of their aan-Dative examples in the database are metaphorical
uses of the kind illustrated in (15b), where an abstract connection is estab-
lished.
(15) a. De beste mensentredmolens zijn uitgerust met een hartslagmeter die de
draver met een draadje aan zichzelf verbindt. [NRC]
‘The best human treadmills are equipped with a heart rate monitor, which
the runner connects to himself by means of a wire.’
b. Verder heeft Ullrich zijn naam verbonden aan een juniorenteam van
Telekom. [NRC]
‘Furthermore, Ullrich has linked his name with Telekom’s junior team.’
With all other verbs in Table 2, the established ‘contact’ relation can only
be metaphorical. The use of geven ‘give’ and other verbs of possessional
transfer in the construction, for instance, seems to be motivated by a ‘Pos-
session is contact’ metaphor. One of the characteristics of prototypical pos-
session listed in Taylor (1992: 202-205) is that the possessor and the pos-
sessed entity are in a relation of close spatial proximity (so that the
possessor can exercise his rights over the possessed). If possession typically
implies spatial proximity, which in itself is a generalized form of spatial
contiguity or contact, the causation of possession can be seen as the estab-
lishment of a contact relation between two entities: the agent relocates the
theme to a position where it is “in contact” with a new possessor. Note that
this localist account suggests that, diachronically, the use of the aan-Dative
with verbs of giving and closely related classes very probably originated
with prototypical ‘give’ events which involve a physical transfer of a con-
crete entity (i.e., where the theme is actually moved from a position in the
spatial proximity of the agent to a position in the spatial proximity of the
recipient) and has subsequently spread to more abstract ‘transfer’ events.
290 Timothy Colleman

Testing this hypothesis is one of the major tasks for a diachronic study of
the Dutch aan-Dative. In any event, in the present-day language, the aan-
Dative covers a wide variety of ‘caused reception’ events, including events
of possessional transfer which involve some kind of abstract commodity
rather than a concrete object (see e.g. verlenen ‘grant’ and toekennen
‘award’, ranked 25th and 27th in Table 2, respectively, both of which are
verbs of giving which select abstract rather than concrete direct object re-
ferents, such as rights, permits, favours, honours, benefits, etc.) as well as
events of projected rather than actual transfers (e.g. with bieden ‘offer’,
which ranks 24th in Table 2 and gunnen ‘not begrudge’ which is not in-
cluded among the top collexemes in Table 2 but is still among the signifi-
cantly attracted verbs), events of communicative transfer (e.g. with vragen
‘ask’, the 10th strongest collexeme, or with uitleggen ‘explain’, vertellen
‘tell’, voorstellen ‘suggest’, etc., all of which are significantly attracted to
the construction as well), and so on. In fact, the PP of the aan-Dative can
encode very much the same array of (possibly projected, possibly meta-
phorical) recipient/possessor/addressee roles as the bare NP indirect object
of the DOC, which is of course why these two constructions have come to
be thought of as grammatical alternants in the first place.
As is evident from the remaining verbs in Table 2, the aan-Dative cov-
ers a much wider semantic range, though. Another class which is closely
associated with the construction consists of verbs such as besteden ‘devote,
spend’, wijden ‘dedicate, devote’, uitgeven ‘spend’, spenderen ‘spend’,
verspillen ‘waste on’, opofferen ‘sacrifice’, etc. The first two of these are
among the top collexemes of the construction (ranked 3d and 8th, respec-
tively, in Table 2), the other verbs are not listed in Table 2 but follow fur-
ther down the line in the ranking of significant collexemes. As shown by
the instances in (16), aan-Dative clauses with these verbs are fairly close in
meaning to aan-Datives with verbs of giving.
(16) a. Veertig miljoen dollar heeft hij tot nog toe besteed aan zijn onderzoek
naar een reeks Clinton-schandalen. [NRC]
‘Until now, he has spent 40 million dollars on his research into a series of
Clinton scandals.’
b. Ik kan niet de helft van mijn tijd wijden aan het onderhouden van
sociale contacten. [GvA]
‘I cannot devote half of my time to the maintaining of social contacts.’
Typically, the direct object in clauses with these verbs refers to certain re-
sources of the subject referent (time, money, attention, energy, etc.). Just
 The semantics of the Dutch aan-Dative 291

like in clauses with verbs of giving, the aan-object encodes the metaphori-
cal destination of the “goods” which leave the agent’s domain. The differ-
ence with verbs of giving is that in the case of besteden ‘devote, spend’
etc., the goods do not enter into a new relation which is comparable to the
relation they entertain with the agent participant at the onset of the transfer
(i.e., some kind of ‘possession’ relation). Relevantly, despite their partial
similarity with verbs of giving, verbs of the besteden class do not partake in
the dative alternation, which shows that the DOC is much more heavily
constrained when it comes to the range of relations allowed between the
object referents.
The final class to be discussed in this subsection are verbs such as on-
derwerpen ‘subject’ (12th in Table 2) and blootstellen ‘expose’ (not listed in
Table 2, but still among the significantly attracted collexemes), see (17) for
relevant examples.
(17) a. [Hij] veroordeelde ook “kapitalistisch neoliberalisme, dat de menseli-
jke persoon aan blinde marktkrachten onderwerpt”. [NRC]
‘He also denounced “capitalist neoliberalism, which subjects the human
individual to mindless market forces.”’
b. De doorsnee skiër sport niet of veel te weinig. Onvoorbereid stelt hij
zijn lichaam bloot aan de specifieke eisen van de skisport. [GvA]
‘The average skier does not exercise, or exercises way too little. Unpre-
pared, he exposes his body to the specific demands of skiing.’
Such cases represent still another kind of metaphorical ‘caused contact’: the
agent brings the theme within the sphere of influence of the aan-object, has
the theme undergo the working or effects of the aan-object referent (a rela-
tion which, again, implies a certain proximity). More broadly construed,
this class also includes aanpassen ‘adapt’ (22nd in the ranking in Table 2),
as in (18), where the theme undergoes the effects of the aan-object referent
as well, in the sense of being changed to better match it.
(18) De wijziging van de grondwet [is] nodig om het briefgeheim aan te pas-
sen aan nieuwe elektronische ontwikkelingen. [NRC]
‘The revision of the constitution is necessary to adapt the confidentiality
of the mail to new electronic developments.’
Summing up, it is clear that the aan-Dative is home to wide semantic vari-
ation in the ‘contact’ dimension: in addition to actual spatial contact, the
construction can encode the establishment of all kinds of other, more ab-
stract relations between the object referents. Of course, while the discussed
examples illustrate some of the most prominent uses with several represen-
292 Timothy Colleman

tatives among the top collexemes, they do not exhaust the semantic possi-
bilities in this department: aan-Dative clauses with other (semantic classes
8
of) verbs may denote still other forms of ‘caused contact’.

4.4. Semantic variation with regard to the ‘direction’ dimension

The examples in the previous subsection differ in terms of the exact nature
of the ‘contact’ relation involved, but they are crucially alike in that the
subject referent causes the establishment of a relation between the object
referents (the agent brings the theme into contact with the aan-object, so to
speak). The denoted event need not proceed in that direction, however. In
Section 4.1 above, I briefly discussed an example in which the Dutch
Double Object Construction encodes a possessional transfer away from
rather than towards the indirect object referent (see ex. 13). Newman
(1996: 115-118) comments on the tendency observed in several languages
to mark the (human) source in an act of taking in the same way – i.e., by
means of the same case or adposition – as the recipient in an act of giving,
citing examples from German, Czech, Japanese and Chamorro. In such
cases, he argues, the source meaning is never the core meaning of the case
or adposition in question, but represents a semantic extension which is mo-
tivated by the fact that possessional sources are in certain respects quite
similar to recipients (both are non-agentive participants who are not di-
rectly manipulated but experience a change in the state of affairs because of
the agent’s action). In languages like Dutch and English, sources are gener-
ally not marked in the same way as recipients (but by means of specialized
prepositions such as from and its Dutch equivalent van). There are excep-
tions, however: next to all kinds of verbs of (projected) giving, the Double
Object Constructions of both languages also accommodate a small number
of verbs which denote events with the reverse directionality. Goldberg
(2002) posits the constructional subsense ‘X causes Y to lose Z’ for the
English Double Object Construction to account for clauses such as Mina
cost Mel his job. In response to the possible criticism that such uses are too
far removed from the prototypical ‘caused reception’ sense to support the
claim of a unified construction, she argues that the concepts of giving and
taking away are related through antonymy:
[I]t is clear that both the negation and the antonym of a particular concept
are closely associated with that concept. For example, a concept and its
antonym typically serve as strong associates for one another in psycholin-
 The semantics of the Dutch aan-Dative 293

guistic studies (Meyer and Schvaneveldt 1971): e.g., hot primes cold, high
primes low, and giving primes taking away … In this way, we can see that
giving, not giving, and taking away are in fact closely associated concepts.
(Goldberg 2002: 333)
In Dutch, not only the Double Object Construction, but also the aan-
Dative can be combined with a small number of verbs of taking away, as
illustrated in (19) below. (19a) involves the same verb as the double object
example in (13) above, viz. ontnemen ‘take away’, (19b) involves another
prefixed verb with ont- ‘away’, viz. onttrekken ‘withdraw, extract from’,
9
and (19c) is an example with kosten, the Dutch equivalent of English cost.
(19) a. Het Amerikaanse anti-communisme begon op 7 november 1917, de
dag dat de bolsjewieken in Rusland de macht ontnamen aan de sociaal-
democraten. [Sta]
‘American anti-communism began on November 7th 1917, the day when
the Bolshevists took the power away from the social democrats in Rus-
sia.’
b. “We weten allemaal dat Iljoemzjinov corrupt is”, zei Short. “Maar hij
geeft in elk geval geld aan de Fide. Zijn voorganger Campomanes was
corrupt en onttrok geld aan de Fide.” [NRC]
‘“We all know that Iljoemzjinov is corrupt”, Short said. “But at least he
gives money to FIDE. His predecessor Campomanes was corrupt and ex-
tracted money from FIDE.”’
c. … de sociale begeleidingsmaatregelen die aan de Belgische staat 2,1
miljard kosten. [GvA]
‘… the social support measures which cost the Belgian state 2,1 billion
francs.’
All three examples denote events in which the referent of the aan-object
loses possession of the direct object referent. If possession is metaphorical
contact, then the agent does not act to establish a contact relation between
the object referents in these cases, but rather to cancel such a relation. None
of these verbs of dispossession is attested sufficiently often in the aan-
Dative in the corpus to qualify among the top collexemes in Table 2.
However, there are a number of other verbs in Table 2 which are rel-
evantly similar to the verbs in (19), in that they, too, denote events with
“reverse” directionality, i.e. in which the theme originates with the aan-
object referent rather than being brought into contact with it. Good exam-
ples include danken ‘owe to, have to thank for’ (5th in Table 2) and ontle-
nen ‘borrow, derive’ (29th in Table 2). In the examples in (20), the aan-
294 Timothy Colleman

object referent is the source of the relation between the subject and direct
object referents (the difference with the examples in 19 being that it is not a
possessional source). Overhouden ‘keep, be left with’ (14th in Table 2) fits
in this category as well.
(20) a. Die wetenschap dank ik aan de Nieuwe Revu. [NRC]
‘I owe that knowledge to the Nieuwe Revu [a Dutch periodical].’
b. Amerika ontleent veel zelfrespect en kracht aan zijn nationale trots.
[NRC]
‘America derives much self-esteem and strength from its national pride.’
From here it is but a small step to the examples in (21), with wijten ‘blame,
ascribe’ (6th in Table 2) and toeschrijven ‘attribute, ascribe’ (11th in Table
2). In such examples, the aan-object refers to the entity, circumstance or
event which the subject referent holds responsible for the attribute or the
event referred to by the direct object. In other words, the aan-object en-
codes the cause of the direct object referent, which is another kind of meta-
phorical source (see, e.g., Lakoff and Johnson 1999: Ch. 11 for discussion
10
of the general ‘Causes are sources’ metaphor of causation).
(21) a. Bestuurster J. Smeets van de Vervoersbond FNV wijt de onrust die nu
ontstaan is aan de KLM-directie zelf. [NRC]
The director of the Transport Union FNV, J. Smeets, blames the commo-
tion that has arisen now on the KLM-Management themselves.’
b. Tonny de Jong werd zestiende en schreef haar slechte prestatie toe aan
maagklachten. [NRC]
‘Tony de Jong finished sixteenth and attributed her bad performance to
stomach complaints.’
In sum, the aan-Dative displays various uses which differ from the posited
‘caused contact’ prototype with regard to the direction of the denoted
event, i.e. where the theme originates with the aan-object referent rather
than being brought into contact with it.
These source uses are interesting from a diachronic perspective as well.
In the preliminary studies of the diachrony of the dative alternation by Wei-
jnen and Gordijn (1970) and Duinhoven (2003), it is claimed that the pro-
ductive use of aan as a marker for the recipient in all kinds of ‘caused pos-
session’ events is a post Middle Dutch innovation, dating from the
language stage traditionally known as Early Modern Dutch (c. 1500-1650).
As these studies are not corpus-based in the modern sense of the word, the
textual evidence presented in support of this claim is rather sketchy in na-
 The semantics of the Dutch aan-Dative 295

ture. However, if we are indeed dealing with a post Middle Dutch phenom-
enon, this would imply that the use of aan to mark the source of a posses-
sional transfer predated its use as a recipient marker by several centuries,
for the lemma ane in the dictionary of Early Middle Dutch (VMNW) cites
several 13th Century examples of the kind illustrated in (22), where the
subject referent clearly is at the recipient end and the referent of the aan-
phrase at the source end of the transfer. Such observations call for extensive
diachronic corpus research into the semantic evolution of the aan-Dative.
(22) a. dit ghelt sal de meester gheuen stappants alse de leerlinc beghint int
ambacht te leerne, ende de meester salt vort an den leerlinc nemen.
(Ghent, 1280)
‘The master will pay this sum immediately when the apprentice begins to
learn the craft, and the master shall then collect it [the money] from the
apprentice.’
b. alle der ander vriheden ende priuilegien die si (…) an ons verworven
hebben. (Dordrecht, 1284)
‘all the other rights and privileges which they have received from us.’

4.5. Prospects for future research

While the above discussion of the verb classes which most typically instan-
tiate the aan-Dative in real language and of the associated constructional
subsenses constitutes a crucial first step towards an adequate account of the
construction’s multidimensional semantic structure, the analysis outlined
on the previous pages needs fleshing out. Virtually all of the top verbs in-
cluded in Table 2 have been discussed in connection with either the ‘con-
tact’ or the ‘direction’ dimension, but closer scrutiny of the full list of 93
significantly attracted collexemes might reveal additional uses, which differ
from the hypothesized ‘caused contact’ prototype along other axes of vari-
ation. Also, while a verb-centered analysis is the obvious first step to be
taken, a detailed analysis of the lexical material inserted in the aan-Dative’s
argument slots will most probably shed additional light on the construc-
tional semantics.
In addition to this fleshing out of the qualitative semantic analysis, there
are several possibilities for extending the present study in a more quantita-
tive direction. For one, although the collexeme data provide some informa-
tion about the relative degree of salience or structural weight of the various
observed uses – it is obvious from the list of top collexemes in Table 2 that
296 Timothy Colleman

uses in which the aan-PP encodes the “destination” of the direct object
referent are more central to the construction’s semantics than uses with the
reverse directionality, for instance – such quantitative aspects of semantic
structure could be investigated in a much more systematic way by hand-
coding a representative sample of instances for their positions along each of
the distinguished dimensions of semantic variation and statistically evaluat-
ing the results.
Another avenue for future research consists in the translation of the
above semantic distinctions into a set of operationalisable semantic vari-
ables with a view to a multifactorial study of the dative alternation in
Dutch. In recent years, several grammatical alternation phenomena have
been studied from a quantitative multifactorial perspective, see e.g. Gries
(2003) on particle placement in English, Heylen (2005) on word order vari-
ation in the German “Mittelfeld”, Grondelaers, Speelman, and Geeraerts
(2008) on presentative sentences with and without er ‘there’ in Dutch and
Szmrecsanyi (2010) on the English genitive alternation. The English dative
alternation has been investigated with the help of advanced multivariate
techniques as well: Bresnan et al. (2007), for instance, present several lo-
gistic regression models of the English dative alternation based on data
from the American Switchboard corpus. These models include up to four-
teen explanatory variables, one of which is the variable SEMANTIC CLASS,
which is operationalized as a distinction between five broad verb classes
(viz. ‘abstract’, ‘transfer of possession’, ‘future transfer of possession’,
‘prevention of possession’ and ‘communication’). Other explanatory vari-
ables include the discourse-accessibility of the theme and recipient partici-
pants, the pronominality of both objects, the length difference between
them, and so on. Although Bresnan and colleagues are able to explain over
90% of the attested variation in this way, the inclusion of a more fine-
grained set of semantic variables along the lines of the multidimensional
analysis sketched in the previous subsections could even further enhance
the predictive value of such models. In tandem with the similar analysis of
the Dutch DOC in Colleman (2009b) and with the Distinctive Collexeme
Analysis in Colleman (2009a), the multidimensional analysis of the aan-
Dative presented here points towards the semantic variables that could be
included in a multifactorial study of the dative alternation in Dutch.
 The semantics of the Dutch aan-Dative 297

5. Conclusion

Existing studies of the Dutch aan-Dative are mainly concerned with the
identification of the schematic semantic contrasts between the aan-Dative
and the DOC. The results from our corpus-based analysis, however, il-
lustrate that the aan-Dative covers a much wider region in semantic space.
Only a subset of the verbs which are revealed to be significantly attracted to
the construction in the investigated newspaper corpus are verbs of giving:
the rest belong to a variety of other verb classes, which do not partake in
the dative alternation and are as a result hardly ever taken into account in
the relevant literature. Some of these additional verb classes bear a quite
close semantic relationship to verbs of giving (e.g. the verbs of the besteden
‘spend, devote’ class), but with others, there is no immediately obvious
semantic link with the ‘possessional transfer’ use of the construction (e.g.
verbs such as danken ‘have to thank for’ and ontlenen ‘derive from’). I
have characterized the aan-Dative’s semantic structure as consisting of a
family of ‘caused contact’ senses, and have posited a number of semantic
extensions along two different dimensions of semantic variation, viz. the
exact nature of the ‘contact’ relation involved and the direction of the de-
noted event (i.e., whether the aan-object codes a destination or a source).
These extensions motivate the aan-Dative uses of several of the verb
classes which are well-represented among the top collexemes of the con-
struction. To end on a methodological note, the above analysis has shown
that the ranking of significantly attracted verbs resulting from a Collexeme
Analysis provides a good starting point for an exploration of the semantic
range of argument structure constructions, as it offers a more intricate view
on the interplay between lexical and grammatical items than the one af-
forded through the observation of raw frequencies alone. The identification
of clusters of semantically similar verbs among the top collexemes provides
a crucial first step towards the elucidation of the multidimensional semantic
structure of the construction under investigation.

Notes

1. The research reported on in this paper was funded by the Special Research
Fund of Ghent University (BOF/GOA project nr. B/05971/01 ‘Meaning in be-
tween structure and the lexicon’, research unit Contragram). I would like to
298 Timothy Colleman

thank the editors of the volume and series as well as an anonymous reviewer
for their helpful com-ments and suggestions. The usual disclaimers apply.
2. The term prepositional dative is widely used to refer to this kind of ‘transfer’
construction (i.e., with the recipient marked by a preposition) in several lan-
guages. To avoid confusion, though, it should be added that this construction
does not involve dative case marking in Dutch (nor in English, for that mat-
ter). While Middle Dutch was a case language not unlike present-day German,
the modern variety of the language does not exhibit morphological case mark-
ing, except for the distinction between the nominative and non-nominative
forms of personal pronouns. Hence, the NP in the complement of aan in (1b)
does not bear case morphology.
3. This is not to say that the use of aan to introduce a prepositional object and its
use in oblique locative phrases are completely unrelated. On the contrary, in
what follows, I will present a semantic analysis of the aan-Dative which cru-
cially refers to the basic spatial relation aan denotes in clauses such as (10).
This implies that, diachronically, the construction in (8) and (9) must have de-
veloped out of the construction in (10).
4. Space prevents a more elaborate discussion, but it can be observed that with
regard to all the traditional distributional tests for prepositional objects dis-
cussed in, e.g., Broekhuis (2004), the aan-phrases in (8) and (9) behave rel-
evantly alike, whereas the aan-phrase in (10a) behaves differently. To give
just one example, the aan-phrase can easily be extraposed to the right periph-
ery in verb-final geven ‘give’ and besteden ‘spend’ clauses, but not in hangen
‘hang’ clauses, as illustrated below.
(i) … dat hij een cadeautje geeft aan de kinderen.
‘… that he gives a present to the children.’
(ii) … dat hij zijn geld besteedt aan zijn postzegelverzameling.
… that he spends his money on his stamp collection.’
(iii) *… dat hij zijn jas hangt aan de kapstok.
‘… that he hangs his coat on the hatrack.’
5. Similar though not identical solutions of the formal demarcation problem pre-
sented by the aan-Dative are to be found in De Schutter (1974) and Schermer-
Vermeer (1991).
6. The codes in brackets in these and all following examples from the ConDiv
corpus refer to the newspapers included in the cor-pus: NRC = NRC Handels-
blad, GvA = Gazet van Antwerpen, Sta = De Standaard. In all quoted corpus
examples, the relevant verb is in bold type.
7. Note that, just like locative aan-adjuncts, such resultative phrases fail the test
for prepositional objecthood mentioned in note 3 above:
(i) * … dat hij de melk brengt aan de kook.
(ii) * … dat hij de vaas sloeg aan diggelen.
 The semantics of the Dutch aan-Dative 299

8. The main text does not discuss the top verb in Table 2, for instance, viz. to-
evoegen ‘add’. The link with ‘caused contact’, however, is quite straightfor-
ward: if somebody adds X to Y, this results in a situation where X and Y are
intermingled or annexed. Toevoegen can be used to denote concrete (e.g. add-
ing milk to coffee) as well as abstract events (e.g. adding a new task to s.o.’s
responsibilities).
9. As it happens, this is one of the areas of semantic contrast between the Dutch
aan-Dative and the English to-Dative, as the use of verbs of taking away in
the to-Dative is downright impossible (a restriction which relates to the alla-
tive semantic import of the preposition to). This Dutch-English contrast is dis-
cussed in more detail in Colleman & De Clerck (2009).
10. Future research will have to test to what extent the high scores of verbs such
as wijten and toeschrijven are a genre effect: it is intuitively plausible that the
high rankings of these verbs in Table 2 are at least partially due to the use of a
newspaper corpus, which, needless to say, contains many statements about
(perceived) causes and effects.

References

Beliën, Maaike
2002 Force dynamics in static prepositions: Dutch aan, op, and tegen. In
Perspectives on Prepositions, Hubert Cuyckens and Günter Radden
(eds.), 195–209. Tübingen: Niemeyer.
Bresnan, Joan, Anna Cueni, Tatiana Nikitina and Harald Baayen
2007 Predicting the Dative alternation. In Cognitive Foundations of Inter-
pretation, Gosse Bouma, Irene Kraemer and Joost Zwarts (eds.), 69–
94. Amsterdam: Royal Netherlands Academy of Science.
Broekhuis, Hans
2004 Het voorzetselvoorwerp [The prepositional object]. Nederlandse
Taalkunde 9: 97–131.
Colleman, Timothy
2006 De Nederlandse datiefalternantie: een constructioneel en corpus-
gebaseerd onderzoek [Dative alternation in Dutch: a constructionist
and corpus-based investigation]. Unpublished Ph.D. dissertation,
Ghent University.
2009a Verb disposition and argument structure alternations: A corpus study
of the Dutch dative alternation. Language Sciences 31: 593–611.
2009b The semantic range of the Dutch Double Object Construction. A
collostructional perspective. Constructions and Frames 1: 190–221.
300 Timothy Colleman

Colleman, Timothy and Bernard De Clerck

2009 ‘Caused Motion’? The semantics of the English to-dative and the
Dutch aan-dative. Cognitive Linguistics 20: 5–42.
Cuyckens, Hubert
1991 The Semantics of Spatial Prepositions in Dutch. A Cognitive-
Linguistic exercise. Unpublished Ph.D. dissertation, University of
Antwerp.
Daelemans, Walter, Jakob Zavrel and Peter Berck
1996 Part-of-speech tagging for Dutch with MBT, a memory-based tagger
generator. In Congresboek van de Interdisciplinaire Onderzoekscon-
ferentie Informatiewetenchap 1996. Delft: TU Delft.
Davidse, Kristin
1996 Functional dimensions of the dative in English. In The Dative. Vol.
1. Descriptive studies, William Van Belle and Willy Van Langen-
donck (eds.), 289–338. Amsterdam / Philadelphia: John Benjamins.
De Schutter, Georges
1974 De Nederlandse zin: Poging tot beschrijving van zijn structuur [The
Dutch clause: An attempt at describing its structure]. Brugge: De
Tempel.
Duinhoven, Anton
2003 Actieve en passieve indirect objecten. Een kwestie van analyse en
synthese [Active and passive indirect objects. A question of analysis
and synthesis]. Nederlandse Taalkunde 8: 199–230.
Geeraerts, Dirk
1998 The semantic structure of the indirect object in Dutch. In The Dative.
Vol. 2. Theoretical and contrastive studies, Willy Van Langendonck
and William Van Belle (eds.), 185–210. Amsterdam / Philadelphia:
John Benjamins,.
Givón, Talmy
1984 Direct object and dative shifting: Semantic and pragmatic case. In
Objects. Towards a theory of grammatical relations, Frans Plank
(ed.), 151–182. New York: Academic Press.
Goldberg, Adele E.
1995 Constructions: A Construction Grammar approach to argument
structure. Chicago: University of Chicago Press.
2002 Surface Generalizations: An alternative to alternations. Cognitive
Linguistics 13: 327–356.
Goldberg, Adele E., Devin Casenhiser and Nithya Sethuraman
2004 Learning argument structure generalizations. Cognitive Linguistics
14: 289–316.
 The semantics of the Dutch aan-Dative 301

Green, Georgia M.
1974 Semantics and Syntactic Regularity. Bloomington: Indiana Univer-
sity Press.
Gries, Stefan Th.
2003 Multifactorial Analysis in Corpus Linguistics: A study of particle
placement. London / New York: Continuum Press.
2004 Coll.analysis 3.0, R-script for collostructional analysis.
Gries, Stefan Th., Beate Hampe and Doris Schönefeld
2005 Converging evidence: Bringing together experimental and corpus
data on the association of verbs and constructions. Cognitive Lin-
guistics 16: 635–676.
Gries, Stefan Th. and Anatol Stefanowitsch
2004a Extending Collostructional Analysis: A corpus-based perspective on
“alternations”. International Journal of Corpus Linguistics 9: 97–
129.
2004b Co-varying collexemes in the into-causative. Language, Culture,
and Mind, Michel Achard and Suzanne Kemmer (eds.), 225–236.
Stanford: CSLI.
Grondelaers, Stefan, Katrien Deygers, Hilde van Aken, Vicky Van den Heede and
Dirk Speelman
2000 Het ConDiv-corpus geschreven Nederlands [The ConDiv-corpus of
written Dutch]. Nederlandse Taalkunde 5: 356–63.
Grondelaers, Stefan, Dirk Speelman and Dirk Geeraerts
2008 National variation in the use of er ‘there’. Regional and diachronic
constraints on cognitive explanations. In Cognitive Sociolinguistics:
Language variation, cultural models, social Systems, Gitte Kristian-
sen and Rene Dirven (eds.), 153–204. Berlin / New York: Mouton
de Gruyter.
Harley, Heidi
2002 Possession and the Double Object Construction. Yearbook of Lin-
guistic Variation 2: 29–68.
Heylen, Kris
2005 A quantitative corpus study of German word order variation. In
Linguistic Evidence: Empirical, theoretical and computational per-
spectives, Stephan Kepser and Marga Reis (eds.), 241–264. Berlin /
New York: Mouton de Gruyter.
Herriman, Jennifer
1995 The Indirect Object in Present-Day English. Göteborg: Acta Univer-
sitatis Gothoburgensis.
Hilpert, Martin
2008 Germanic Future Constructions. A usage-based approach to lan-
guage change. Amsterdam / Philadelphia: John Benjamins.
302 Timothy Colleman

this vol. The force dynamics of English complement clauses: A collostruc-

tional analysis.
Huddleston, Rodney and Geoffrey K. Pullum
2002 The Cambridge Grammar of the English Language. Cambridge:
Cambridge University Press.
Janssen, Theo
1976 Giving in Dutch: An intra-lexematical and inter-lexematical descrip-
tion. In The Linguistics of Giving, John Newman (ed.), 267–306.
Amsterdam / Philadelphia: John Benjamins.
Lakoff, George and Mark Johnson
1999 Philosophy in the Flesh. New York: Basic Books.
Langacker, Ronald W.
1991 Concept, Image, and Symbol: The cognitive basis of grammar. Ber-
lin / New York: Mouton de Gruyter.
2000 A dynamic usage-based model. In Usage Based Models of Lan-
guage, Michael Barlow and Suzanne Kemmer (eds.), 1–63. Stan-
ford: CSLI.
Meyer, David and Roger Schvaneveldt
1971 Facilitation in recognizing pairs of words. Journal of Experimental
Psychology 90: 227–234.
Newman, John
1996 Give. A cognitive-linguistic study. Berlin / New York: Mouton de
Gruyter.
Panther, Klaus-Uwe
1997 Dative alternation from a cognitive perspective. In Human Contact
through Language and Linguistics, Birgit Smieja and Meike Tasch
(eds.), 107–126. Frankfurt/Main: Lang.
Pinker, Steven
1989 Learnability and Cognition: The acquisition of argument structure.
Cambridge, Mass.: MIT Press.
Schermer-Vermeer, Ina
1991 Substantiële versus Formele Taalbeschrijving: Het indirect object in
het Nederlands [Substantial versus formal language analysis: The
indirect object in Dutch]. Amsterdam: Universiteit van Amsterdam.
2001 Grammatica, lexicon en de dubbelobject-constructie in het Neder-
lands en het Engels [Grammar, the lexicon and the Double Object
Construction in Dutch and English]. Nederlandse Taalkunde 6: 22–
37.
Speelman, Dirk
1997 Abundantia Verborum. A computer tool for carrying out corpus-
based linguistic case studies. Unpublished Ph.D dissertation, Catho-
lic University of Leuven.
 The semantics of the Dutch aan-Dative 303

Stefanowitsch, Anatol and Stefan Th. Gries

2003 Collostructions: Investigating the interaction of words and construc-
tions. International Journal of Corpus Linguistics 8: 209–43.
Szmrecsanyi, Benedikt
2010 The English genitive alternation in a cognitive sociolinguistics per-
spective. In Advances in Cognitive Sociolinguistics, Dirk Geeraerts,
Gitte Kristiansen and Yves Peirsman (eds.), 141–166. Berlin / New
York: Mouton de Gruyter.
Talmy, Leonard
2000 Toward a Cognitive Semantics. Cambridge: MIT Press.
Taylor, John
1992 Linguistic Categorization. Oxford: Clarendon Press.
Van Belle, William and Willy Van Langendonck
1996 The indirect object in Dutch. In The Dative. Vol. 1. Descriptive
studies, William Van Belle and Willy Van Langendonck (eds.),
217–250. Amsterdam / Philadelphia: John Benjamins.
Van Langendonck, Willy
2000 Semantic role configurations and syntactic patterns of the Dutch
indirect object. In The Berkeley Conference on Dutch Linguistics
1997, Thomas Shannon and Johan Snapper (eds.), 79–92. Lanham:
UPA.
VMNW = Pijnenburg, Willy, Karina Van Dalen-Oskam, Katrien Depuydt and
Tanneke Schoonheim
2001 Vroegmiddelnederlands Woordenboek [Early Middle Dutch Dic-
tionary]. Leiden: Gopher.
Weijnen, Antonius A. and Jaap Gordijn
1970 Het binnendringen van het voorzetsel in het indirekt objekt in het
Nederlands [The entrance of the preposition in the indirect object in
Dutch]. Leuvense Bijdragen 59: 125–130.
Wierzbicka, Anna
1986 The semantics of ‘internal dative’ in English. Quaderni di Semantica
7: 121–35.
Wulff, Stefanie
2006 Go-V vs. go-and-V in English: A case of constructional synonymy?
In Corpora in Cognitive Linguistics. Corpus-based approaches to
syntax and lexis, Stefan Th. Gries and Anatol Stefanowitsch (eds.),
101–126. Berlin / New York: Mouton de Gruyter.
Zavrel, Jakob and Walter Daelemans
1999 Recent advances in memory-based part-of-speech tagging. In Actas
del VI Simposio Internacional de Comunicacion Social, Centro de
Lingüística Aplicada (ed.), 590–597. Santiago de Cuba: Ediciones
Editorial Oriente.
Corpus-based evidence for
an idiosyncratic aspect-modality relation in Russian

Dagmar Divjak

Abstract
There is an abundance of literature suggesting a relationship between aspect and
modality; typically, perfective aspect is related to “objective” or “factive” informa-
tion, whereas imperfective aspect is linked to “subjective”, “perspectivized” or
1
“counterfactual” information (Boogaart and Janssen 2007: 821). More specifically,
the typological linguistic literature posits that imperfective infinitives are used to
express epistemic modality while perfective infinitives render deontic meanings
(Abraham 1991; Abraham and Leiss 2006). This position is strongly disputed by
Slavic linguists, however, and by Russian linguists in particular (Padučeva 2006;
Šmelev and Zalizniak 2006; Wiemer ms.).
In this paper, I will present empirical evidence supporting the stance that in
Russian, imperfective infinitives typically render deontic meanings, whereas per-
fective infinitives preferably express dynamic modality. The focus is on presenting
how a corpus-based study was instrumental in answering this question and how the
data collected were annotated and analyzed statistically to make drawing reliable
conclusions possible. I will go on to interpreting the results from a theoretical,
cognitive linguistic perspective, and I will argue that the “reverse outcome” is to be
expected on the basis of what we know about the semantics of the Russian aspec-
tual system and general principles of cognition.
Keywords: aspect, modality, Russian, usage-based, mixed effects modeling

1. On time

Time is a fundamental aspect of our existence. It is crucial for motor con-

trol in walking, talking, playing sports and so many other daily activities
that we perform ‘automatically’. We all agree time exists, although our
sense of time is not associated with a specific sensory system. Instead, it is
the brain that governs time perception and keeps time by means of rhythms,
ranging from neural (0.001 to 10 seconds) over circadian (24 hours) to
ovarian (28 days) (Goldbeter 2008). Traditionally, the way in which time is
perceived, represented and estimated has been explained using a pace-
306 Dagmar Divjak

maker–accumulator model. However, recent advances have challenged this

traditional view. It is now proposed that the brain represents time in a dis-
tributed manner and tells the time by detecting the coincidental activation
of different neural populations (Buhusi and Meck 2005). We are largely
unaware of this “human timing machine”. For us, time is rather an experi-
ence that emerges through the ongoing interaction between our bodies and
the surrounding environment.
Yet, time is not restricted to physical experience. Almost everything we
talk about involves time and talking itself takes place in time. And probably
because we experience time as so important, we have developed a myriad
of ways to express time in language. In many languages, the very fact that
the situation and the act of talking about that situation are temporally re-
lated (and ordered) is captured by tense(s): speech time precedes, coincides
with or follows event time. Yet, tense is merely one way of dealing with
time in language, albeit the primary one in Germanic and Romance lan-
guages. Slavic languages, for example, assign equal (if not more, see Janda
2002: 6) importance to the fact that time seems to affect different situations
differently. Some situations can go on and on, such as live, while others
come to their end reasonably quickly, think of be born and die. Other situa-
tions, still, happen too quickly for us to observe them properly, such as
blink or sneeze. These, and many more distinctions, are expressed in Slavic
(and other) languages by means of aspect(s), different ways of viewing the
internal temporal constituency of a situation (Comrie 1976: 3).
Some languages feature both tense and aspect. The Slavic languages for
instance, including Russian, do. Whereas the Russian tense system is rather
simple (past, present, future), aspect has reached an unusually high degree
of grammaticalization. Although in Russian “only” two aspects exist, im-
perfective and perfective, aspect is, for most verbs, an obligatory morpho-
logical feature. This has not always been the case: Bermel (1997: 476)
shows how “the history of Russian aspect is the retreat of a lexically based
aspectual system in favour of a more grammatically and contextually based
one”. As a consequence of this grammaticalization process, at present, there
is no abstract notion of a “pure” situation for speakers of Russian: all situa-
tions come “pre-packaged” with an imperfective or perfective aspectual
designation (Janda 2002: 5, 7). In other words, the conceptualization of
situations as they are set in grammar unavoidably contains properties of
how time affects a situation.
In this paper, I will look into how situations, prepackaged by grammar
according to the way in which they react to time, behave in one particular
 Idiosyncratic aspect­modality relation in Russian 307

context, a modal one. More in particular, I will be looking at lexical expres-

sions of modality as captured by predicative adverbials that combine with
an infinitive. Incidentally, these two domains, aspect and modality, repre-
sent the two extremes of the semantic cline from abstract grammar to con-
crete lexicon, respectively. As such, the results of this study will also yield
valuable insights into the functioning of the grammar-lexis interface. Con-
ceptual structure, be it grammatical or lexical in nature, is said to be moti-
vated by usage. Therefore, the main part of this paper deals with the ins and
outs of a corpus-based investigation into the interaction of modality and
aspect. Although the findings reverse the claims made in the literature, I
will show that the way in which aspect and modality line up in Russian can
be accounted for on the basis of what we know about Russian aspect and by
referring to the general cognitive principles that organize language.
After an introduction to the dispute in Section 2, I will zoom in on the
two central concepts, aspect and modality in Section 3, and provide an op-
erational definition of these concepts that is suited to serve as a basis for
corpus annotation. In Section 4, I will briefly describe how the corpus was
compiled as well as how the data were annotated. Next, in Section 5, a
mixed effects logistic regression model will be fit to the data. Finally, in
Section 6, the results will be interpreted and evaluated from a cognitive-
linguistic point of view.

2. The dispute

Much effort has been put into clarifying the relation between modality and
other verbal properties, in particular mood and tense. Until recently, the
relation between modality and aspect received much less attention, how-
ever (cf. Boogaart and Janssen 2007: 821). For Russian this situation is
particularly unfortunate: in Russian, aspect has grammaticalized to an un-
usual extent, hence the imperfective versus perfective distinction is obliga-
torily marked on all verbal forms (Dahl 1985: 74–85). The assumption
underlying much of the research into the interaction between aspect and
modality expects imperfective aspect to prevail in modal constructions or to
be used to express epistemic or alethic modality, whereas perfective aspect
renders deontic meanings. This hypothesis does not seem to hold for Rus-
sian, however (cf. Wiemer ms.).
Surveying the literature reveals some remarkably conflicting statements
with respect ot the interaction between aspect and modality. It has become
308 Dagmar Divjak

customary to assert that imperfective aspect, by and large, prevails in modal

constructions (see Trnavac 2006: 1–9 for an overview). Yet, for Russian, it
has traditionally been claimed that perfective aspect prevails in modal con-
structions rendering non-repetitive actions (Rassudova 1968: 55), or at least
that there is no clear tendency to use one aspect to the exclusion of the
other in modal constructions expressing necessity and obligation (Forsyth
1970: 263). In addition, the general linguistic literature posits that imper-
fective infinitives are used to express epistemic modality while perfective
infinitives render deontic meanings (Abraham 1991, Abraham and Leiss
2006). This claim is contested by Slavic linguists who state that imperfec-
tive aspect expresses deontic meanings while perfective aspect renders
dynamic modality (Padučeva 2006, Šmelev and Zalizniak 2006, Wiemer
ms., to name but the participants of the SLE 2006 workshop). The two sen-
tences used by Šmelev and Zalizniak (2006) to dispute Abraham and Leiss’
(2006) claims are the following:
(1) Здесь можно переходить улицу.
Here.Adv possible/permissible.PredAdv cross.impf.inf street.Acc.F.Sg
‘You are allowed to cross the street here.’
(2) Здесь можно перейти улицу.
Here.Adv possible/permissible.PredAdv cross.pf.inf street.Acc.F.Sg
‘You can get across the street here.’
These two sentences only differ in the aspect of the infinitive: in example
(1) an imperfective infinitive is used, whereas example (2) features a per-
fective infinitive. As a consequence, example (1) would render deontic
permissibility, while (2) expresses dynamic possibility.
Recently, Padučeva (2006) put forward the claim that the aspectual form
of the infinitive under negation would correlate with the concepts of possi-
bility and necessity. On this account, non-deontic possibility would be sig-
nalled by means of perfective infinitives in (both positive and) negated
contexts, whereas non-deontic necessity would require imperfective infini-
tives, yet only under negation. Deontic modality would always come with
imperfective infinitives, irrespective of the polarity exhibited by the utter-
ance, just as epistemic modality would go hand in hand with perfective
infinitives. Having confronted her intuition-based account with corpus data,
Padučeva is forced to admit that only negated deontic modality constrains
the aspectual options the infinitive has as predicted and limits it to the im-
2
perfective aspect.
 Idiosyncratic aspect­modality relation in Russian 309

Padučeva’s (2006) work illustrates that introspection-based data, even

when generated by a native speaker-linguist, do not yield reliable results.
Instead of relying on an introspective analysis of decontextualized senten-
ces, I take a strictly corpus-based, quantitative approach to the analysis of
the interaction between aspect and modality. This will make it possible to
provide a verifiable answer to the question of which of the existing hypoth-
eses about the relation between aspect and modality in Russian is correct.
The constructions I focus on are positive and negative declarative sentences
built around a predicative modal adverbial followed by an infinitive, as
illustrated in examples (1) and (2) above.

3. Crucial concepts

The interaction studied in this paper holds between two of the most elusive
concepts in linguistics, aspect and modality. For that reason, I will start by
providing a working definition that will also serve as the basis for the oper-
ationalization of both notions in Section 4.

3.1. Aspect

Comrie’s (1976: 3–5) characterization of aspect as “different ways of view-

ing the internal temporal constituency of a situation” has become the tradi-
tional definition. In using aspect, the language user thus indicates whether a
situation is construed as either bounded or unbounded. Cross-linguistically,
the perfective:imperfective opposition is the most typical “grammatical”
embodiment of this conceptual distinction. Yet, even in languages that have
this central aspectual distinction, it is often absent from non-indicative or
non-finite clauses and perfective aspect may be restricted to past time refer-
ence only.
Russian, like the other Slavic languages, deviates from this cross-
linguistic aspectual norm (Dahl 1985: 74–85) in that aspect in Russian has
grammaticalized to an unusual extent: Russian marks the imperfec-
tive:perfective distinction morphologically on all verbal forms. This obliga-
tory marking of aspect on every verbal form in any time reference makes
Russian particularly suitable for a study of the relationship between aspect
and modality.
310 Dagmar Divjak

For the non-Slavicist, I will provide a rudimental sketch of the imperfec-

tive:perfective distinction in Russian, based on Maslov (1959) and Rassu-
dova (1968: 8). Obviously, many more holistic descriptions have been pro-
vided in more recent times, yet with the exception of Šmelev and Zaliznjak
(2000) and Janda (2003), few have attempted to formulate their proposals
so clearly as to be of use for the language learner or the corpus annotator.
Moreover, many of the more recent proposals mainly look for an invariant
meaning of the imperfective or perfective aspect as a whole, and thus
merely provide synonyms for the two existing grammatical categories. And
finally, contemporary models still aim to account for the meanings distin-
guished five decades ago by Maslov (1959) and Rassudova (1968: 8).
According to Maslov (1959) and Rassudova (1968), the Russian imper-
fective expresses three core meanings:
1. General-factual or simple denotative meaning that merely states or
asks whether or not an event has taken place
Вы читали Войну и Мир?
“Have you read War and Peace?”
2. Processual meaning, focusing on the process of the event taking
place itself
Молодая женшина сидела около окна вагона и читала.
“The young woman was sitting by the window of the carriage and
was reading.”
3. Repetitive meaning, explicitly rendering the repetition of an event
Иногда я перечитывал писателей, которых особенно любил.
“Sometimes I re-read writers whom I really liked.”

The perfective then renders:

1. Concrete/specific-factual meaning, focusing on a particular in-

stance of a situation in its entirety
Он повторил мне свой вопрос.
“He repeated his question for me.”
2. Summarizing meaning, reporting on the whole of a series of re-
peated events as if it were one
Он несколько раз повторил свой вопрос.
“He repeated his question several times.”
 Idiosyncratic aspect­modality relation in Russian 311

3. Visual-exemplary or graphic-illustrative meaning that illustrates a

typical event on the basis of one instance of it
Если вы не поймете мое объяснение, я всегда повторю его вам.
“If you do not understand my explanation, I can/will always repeat
it for you.”

As will be clear from even this brief exemplification, the three core-
meanings distinguished for each aspectual type remind of Structuralist bi-
nary oppositions. Yet, within each aspectual type, the three sub-meanings
resemble a small cognitive radial network. We will return to this property
of aspectual semantics in Section 6.
The Russian aspectual system has three peculiar properties, however,
that need to be explicitly addressed in a corpus-based study: these are the
concepts of ‘aspectual pairs’, (im)perfectiva tanta and biaspectual verbs.
The vast majority of Russian verbs exist in two forms – a perfective form
and an imperfective form. There is a strong tendency to regard these imper-
fective and perfective ‘partners’ as forming an aspectual pair (for an excep-
tion see Janda 2007). Yet, two verbs form a pair if and only if the perfective
merely adds the idea of completion and/or result to the imperfective. An
example here would be pisat/napisat’, imperfective write and perfective
write-ready. These verbs differ from 1) inchoative verbs such as pla-
kat/zaplakat’ where the perfective begin to cry adds the idea of beginning
to the imperfective cry, 2) delimitative verbs where the perfective puts a
limit on the time or effort spent carrying out an action, such as rabo-
tat’/porarbotat’, that translates as work/work for a while and from 3) se-
melfactives that stress the fact that the situation occurred only once, e.g.
prygat’/prygnut’ or jump versus take a jump, jump once.
Despite the prevalence of paired verb forms, some verbs only exist in
the imperfective and express among other things states, e.g. the imperfec-
tiva tanta byt’ ‘be, exist’ and žit’ ‘live’, whereas perfectiva tanta are limited
to the perfective, e.g. stat’ ‘begin, start’ or sostojat’sja ‘take place’. Still
other verbs, the so-called biaspectuals such as atakovat’ ‘attack’, have only
one form that is used in contexts requiring either imperfective or perfective.
The way in which linguists establish these aspectual pairs and the fact that
not all verbs allow aspectual choice will prove to be of importance in Sec-
tion 4.2.
312 Dagmar Divjak

3.2. Modality

Modality is a concept that has been defined in a variety of ways. Roughly,

expressions of modality capture the relation between subject and situation.
A traditional extensional characterization is provided by Bybee, Perkins
and Paglucia (1994: 176). They distinguish between “grams with uses that
are traditionally associated with modality – for instance, those indicating
obligation, probability, and possibility – and those traditionally associated
with mood – imperative, optative, conditional, and subordinate forms”.
Modality defined this way is in turn subdivided, and again this is done dif-
ferently by different researchers. For the purpose of operationalizing the
concept of modality I opted for Nuyts’ (2006: 2–5; see also Nuyts et al.
2005) characterization of modality and will work with two subtypes, i.e.
dynamic and deontic modality.
Dynamic modality (from Greek dynamis, meaning or ‘strength’ or
‘power’) is concerned with the disposition of certain empirical circum-
stances with regard to the occurrence of some event (Perkins 1983: 34).
Nuyts (2006: 2–5) proposes to consider the state of affairs as dynamic if it
stems from a capacity ascribed to the controlling participant in that state of
affairs. This capacity may be participant inherent or participant imposed.
Participant inherent capacities require reference to be made to (participant)
internal factors such as ability or capacity, whereas participant imposed
capacities imply reference to (participant) external factors, e.g. circum-
stances. An example of dynamic modality is given in (3).
(3) Я постучал — никакого ответа. — “Сарториус!” — крикнул я. —
“Доктор Сарториус! Это я, новичок — Кельвин! Мне нужно с вами
увидеться impf.inf, прошу вас, откройте!” [S]
“I knocked – no answer. “Sartorius!” – I yelled. – “Doctor Sartorius! It’s
me, novice Kelvin! I need to see you, please, open up!”
Deontic modality (from Greek deon, meaning “duty” or “that which is
binding”) is a modality relating to normative or valuational classifications
of actions and states of affairs, such as the obligatory, the forbidden, the
permitted. Nuyts (2006: 2–5) suggests considering the source of that neces-
sity, obligation, possibility, permission as deontic if an indication of degree
of moral or social desirability of the state of affairs (henceforth SoA) is
expressed, as is the case in example (4).
 Idiosyncratic aspect­modality relation in Russian 313

(4) И, конечно, совершенно ужасно было бы даже помыслить о том, что

такого человека можно казнить impf.inf. [MM]
“And, of course, just thinking about the fact that such a man could be
executed was so awful.”
Nuyts et al.’s (2005) classification of modality differs from most other
classifications of modal meanings in that it does not (quasi-)automatically
assign deontic status to permissive utterances. Likewise, they question
whether directive meanings are specifically deontic at all. Both permissives
and directives may equally well ensue from a deontic assessment as well as
from observations regarding potential or necessity present in a situation. I
will come back to this issue in Section 4.2.

4. Corpus and variables

After a brief presentation of the corpus consulted for this study (4.1), I will
introduce the variables selected for data annotation (4.2). This will feed into
an exploration of the raw data (5.1) and prepare for confirmatory statistical
analysis (5.2 and 5.3).

4.1. The corpus

It might surprise the reader to see that the data used in this study on the
aspect/modality interaction in one language, Russian, are extracted from a
small Slavic parallel corpus. Yet, a corpus-based quantitative approach
opens up interesting additional avenues of research at a relatively low cost.
This study, for example, was envisaged as the first in a series of studies that
would compare the situation in different Slavic languages (see Divjak and
Dahl 2007, Divjak 2008a and b). Therefore, a major concern was to avoid
spurious results, solely due to differences in corpus structure, yet unrelated
to language or phenomenon studied. Choosing a small parallel corpus over
a large national corpus resolves this issue (see Wälchli 2007 on the advan-
tages of using parallel corpora for typological research). Furthermore,
cross-linguistic comparisons require all examples to be contextually situ-
ated so that authors and translators are more likely to be conceptualizing
exactly the same type of modality and to be selecting the appropriate lin-
guistic means to express it – not a disposable luxury given the volatility of
the concept involved. Again, working on the basis of a Slavic Parallel Cor-
314 Dagmar Divjak

pus and analyzing the data statistically facilitates this (see Wälchli 2007 on
the advantages of using translational equivalents for typological research).
For inclusion in the corpus, highly regarded post-modern literary works
were selected, originally written in a Slavic language, with each language
included in the comparison (Russian, Polish and Bosnian-Croatian-Serbian)
providing one novel (Table 1). Having a Slavic language as source lan-
guage for each of the novels guarantees the availability of high quality
translations in several Slavic languages, while at the same time eliminating
the risk of introducing “translationese” influences from non-Slavic lan-
guages.3
Table 1. Corpus contents
Original Translation Translation
(Russian) Bulgakov, M. (Polish) Mistrz i (Serbian) Мајстор и
1938. Master i Margarita. Małgorzata (by Irena Маргарита (by Milan Čopić)
[MM] Lewandowska and
Witold Dąbrowski)
(Polish) Lem, S. 1961. (Russian) Солярис (by (Serbian) Solaris (by Predrag
Solaris. [S] Dmitrij Bruškin) Obućina)
(Serbian) Pavić, M. 1984. (Polish) Słownik (Russian) Хазарский словарь
Hazarskij Rečnik. [ChH] chazarski (by Elżbieta (by Larisa Savel’evaja)
Kwaśniewska and
Danuta Cirlić-
Straszyńska)

The corpus contains 1 million tokens, with [MM] accounting for roughly
44,4 %, [S] for 22,2 % and [ChR] for 33,3 % of the corpus. It has been
scanned, spell-checked, morpho-syntactically annotated, lemmatized and
aligned and is now included in the Parallel Corpus of Slavic and Other
Languages (ParaSol), a collection of on-line searchable Slavic (post-war
belletristic) texts.
Given that this paper focuses on Russian, the data used in the remainder
of this paper stem from the original Bulgakov as well as from the Russian
translations of Lem’s and Pavić’s work. In the next section I will introduce
the data extracted and the variables selected for annotation.

4.2. The Variables

The point of focus in this paper are the four Russian adverbial predicatives
that combine with an infinitive (see Table 2), i.e. možno (‘it is possible,
 Idiosyncratic aspect­modality relation in Russian 315

permissible’), negated counterpart nel’zja (‘it is impossible, not allowed;

one should not’) for (in)ability, (im)possibility and (non-) permissibility
and (ne) nužno (‘it is necessary; one must, ought, should, needs’), and (ne)
nado (‘it is necessary; one must, ought’) for necessity and obligation. In all,
from the corpus 380 instances were retrieved that contain one of the above
listed modal words and hence qualify for inclusion in this research. Below
are a few examples, one for each modal word:
(5) Да, мы не верим в бога, – чуть улыбнувшись испугу интуриста,
ответил Берлиоз, – но об этом можно говоритьimpf.inf совершенно
свободно. [MM]
“Correct, we do not believe in God, – Berlioz answered, having smiled (a
bit) at the fear of the foreign tourist, – but we can talk absolutely freely
about that.”
(6) Ни в одной лабораторной печи нельзя было получитьpf.inf
температуру, при которой бы они плавились, разве что внутри
атомного котла. [S]
“There was not a single laboratory oven in which a temperature could be
reached at which they would melt, except inside an atomic chamber.”
(7) Выведите его отсюда на минуту, объясните ему, как надо
разговариватьimpf.inf со мной. [MM]
“Take him outside for a minute, and explain to him how (one has) to talk
to me.”
(8) Когда песок переставал шуршать, нужно было перевернутьpf.inf
книгу и продолжать чтение в обратном порядке, с того места, где
остановился, к началу (...) [ChR]
“When the sand stopped rustling, one had to turn the book around and
continue reading in the reverse direction, from the place where you had
stopped to the beginning (...)”
Each of the 380 instances retrieved was manually annotated for the follow-
ing six properties:
(i) the novel it was extracted from (see Table 1)
(ii) the modal word used (see Table 2 below)
(iii) the aspectual range of the infinitive (imperfective only, perfective
only, biaspectual, having both imperfective and perfective forms)
(iv) the aspect of the infinitive (imperfective versus perfective)
(v) the modality type expressed (dynamic versus deontic)
(vi) the polarity of the utterance (positive versus negative).
316 Dagmar Divjak

The first three variables, 1 through 3, have to be considered in order to

be able to model the data correctly. They include the novel the sentence
was extracted from and the modal word that was used; for this particular
study that includes Russian data only, the novel reveals the
author/translator. The aspectual range of the infinitive (imperfective only,
perfective only, biaspectual, having both imperfective and perfective
forms) is included for modelling reasons as well. Infinitives that only exist
in the imperfective or perfective form do not pose a problem for aspect
selection as they will have to appear in that form, independently of the type
of modality that is expressed in the sentence. Precisely because there is no
aspectual choice to be made and the context is central in deciding whether
dynamic or deontic modality is involved, these verbs have to be excluded
from the analysis (see Section 5.2). Verbs that freely occur in both aspects,
however, force the speaker to make the appropriate aspectual choice every
time s/he uses those verbs in a modal context. Yet, the phrase “occurring
freely in both aspects” contains a potential pitfall. Recall that an imperfec-
tive verb is paired up with a perfective verb if the perfective verb adds
nothing but the idea of completion. Clearly, it is tricky to decide in an ob-
jective and verifiable fashion whether or not only the idea of completion
has been added. To ensure objectivity in this matter, every verb was
checked in Ožegov and Švedova (1999): if the verb or verb sense was listed
as having a perfective or imperfective partner, that verb was classified as
such in the data collection. From a semantic point of view, concepts that
form aspectual pairs express situations that are dynamic, durative and telic
on Smith’s (1997) 5-way classification. It should not come as a surprise
that precisely these situations display a particularly interesting behaviour in
modal contexts. The same situations have come up both in historical and
acquisition studies on aspect as bootstraps for the development and acquisi-
tion of the category. Bermel (1997) discovered that non-punctual telic
predicates distinguish aspect regularly in Old Russian already while Stoll
(1998: 372–373) found that aspectual properties of telic verbs are acquired
earliest and easiest by children. One of the reasons behind this, Stoll (1998:
374) suggests, might be that telic predicates render a qualitative change of
state that is salient, even in isolation.
The last three variables, 4 through 6, serve to answer the research ques-
tion and have been identified in literature on related topics as potentially
relevant. Given the hypothesis I set out to test, i.e. that modality type guides
aspectual choice, aspectual choice was encoded as the dependent variable
and modality type (dynamic versus deontic) was included as independent or
 Idiosyncratic aspect­modality relation in Russian 317

explanatory variable. Following Nuyts (2006: 2–5), whose interpretation of

modality I support, I tag a state of affairs as dynamic if it stems from a par-
ticipant inherent or participant imposed capacity ascribed to the controlling
participant in the state of affairs. As example (3) reveals, at this stage no
attempt has been made to distinguish between (participant) internal and
external factors as it has not been suggested in the literature that this dis-
tinction would influence aspectual choice. SoAs are classed as deontic if an
indication of degree of moral or social desirability of the state of affairs is
expressed. Note that all permissives and directives have been judged ac-
cording to this criterion, i.e. have not automatically been given deontic
status. In addition, the polarity of the utterance (positive versus negative)
was taken into account. Polarity has proven to be of importance for assign-
ing aspect in (modal) imperatives as negated imperatives seem to require
imperfective aspect.

5. Data analysis

In this section, I will present raw data on the behaviour of the two inde-
pendent, explanatory variables, i.e. modality type and polarity type, that are
expected to influence the choice of aspect for the infinitive (variable 5).

5.1. The raw data

Recall that 380 sentences were retrieved from the Russian part of the cor-
pus. Out of all 380 available sentences, 314 examples contained an infini-
tive that allows aspectual choice, i.e. exists in both imperfective and perfec-
tive. In Table 2, the four modal words are listed together with the number
of examples encountered in the corpus.
Table 2. Modal adverbial predicatives
Modal Word Instances Retrieved
Можно (‘It Is Possible, Permissible’) 153
Нельзя (‘It Is Impossible, Not Allowed; One Should Not’) 59
Надо (‘It Is Necessary; One Must, Ought’) 41
Нужно (‘It Is Necessary; One Must, Ought, Should, Needs’) 61
Column Total 314

When both variables, modality type and polarity, are cross-tabulated with
aspect (Table 3), we see that examples have been encountered for all vari-
318 Dagmar Divjak

able combinations as far as the imperfective infinitive is concerned; the

ratios between negative and positive dynamic and deontic are in both cases
about 1:2. The perfective shows a less balanced view: the majority of per-
fective infinitives are encountered in positive dynamic sentences, although
negative dynamic sentences can also be constructed using a perfective in-
finitive. Only two examples feature perfective infinitives in positive deontic
sentences and no instances were found in which a perfective infinitive is
used in a negative deontic sentence. This last finding might even qualify as
a rule, and I will come back to it in Section 6.
Table 3. Aspect, modality type and polarity
Aspect of infinitive Modality Type Polarity Row Totals
negative positive
imperfective dynamic 7 15 21
deontic 15 26 41
perfective dynamic 51 198 249
deontic 0 2 2
Column Totals 73 241 314

In other words, the raw data reveal that deontic and negated contexts seem
to prefer imperfective infinitives, whereas dynamic and positive contexts
favour perfective infinitives. Do we have reason to believe that these rela-
tions hold in the Russian literary language in general, or are they just a
fluke found in this particular dataset? And are both variables, modality type
and polarity type, equally good predictors of aspectual choice? What hap-
pens when we combine them?

5.2. Mixed Effects Logistic Regression

Given the quantitative nature of the data collection, answering these ques-
tions requires running confirmatory statistics. As the issue that needs to be
predicted – the choice of imperfective versus perfective aspect – is binary,
the data can be modelled using Logistic Regression Analysis. Logistic re-
gression is a technique used extensively in the medical and social sciences:
for example, the probability that a person has a heart attack within a speci-
fied time period might be predicted from knowledge of the person’s age,
sex and body mass index. Marketing applications also rely on Logistic Re-
 Idiosyncratic aspect­modality relation in Russian 319

gression Analysis for the prediction of a customer’s propensity to purchase

a product or cease a subscription.
All of these outcomes are binary (one lives or dies, one purchases a pro-
duct or doesn’t, one continues or ceases a subscription, a Russian verb is
perfective or imperfective) and a variety of “risk” factors or predictors
(e.g., age, sex, BMI, but also modality type and polarity type) influences
these outcomes. In order to find out how exactly the predictors relate to the
outcome, a logistic model can be fitted. A logistic regression model is used
for predicting the probability of occurrence of a binary event and does so
by fitting data to a logistic curve; in this case, it assesses the likelihood of
encountering an infinitive in a particular aspect in a particular modal con-
text. Logistic Regression is thus a useful way of describing the relationship
between one or more predictors (i.e., modality type, polarity type) and an
outcome such as aspect (which only takes two possible values: imperfective
or perfective).
In order to assess this likelihood, the relation is examined of a binary
dependent variable or response variable (here the imperfective or perfective
aspect the infinitive occurs in), to specified independent variables, a.k.a.
explanatory variables (in our case, these latter variables are modality type
and polarity type). In assessing the relation between dependent and inde-
pendent variables null hypothesis testing plays a major role. The null hy-
pothesis describes (some aspect of) the (statistical) behaviour of a set of
data that is treated as valid unless the actual behaviour of the data contra-
dicts this assumption. Thus, the null hypothesis is contrasted against an
alternative hypothesis. This particular logistic model tests the null hypoth-
esis that the proportion of sentences containing imperfective infinitives is
equal in what is called populations, i.e. deontic versus dynamic modality
and positive versus negative polarity. The alternative hypothesis claims this
would not be the case.
In plain English the null hypothesis for modality type would be phrased
as follows: the probability that an arbitrarily chosen sentence from the dy-
namic population would contain an imperfective infinitive is equal to the
probability that an arbitrarily chosen sentence from the deontic population
would contain an imperfective infinitive. The alternative hypothesis then
states that these two values are not equal in both populations, hence the
probability that an arbitrarily chosen sentence would contain an imperfec-
tive infinitive is a function of which population (deontic or dynamic) the
observation stems from. The same holds for hypothesis about the influence
320 Dagmar Divjak

of polarity on choice of aspect, and any other explanatory variables one

would wish to include.
The results of the analysis are expressed in the form of regression co-
efficients. Each of the regression coefficients describes the size of the con-
tribution of a predictor or regressor. A positive regression coefficient means
that that predictor increases the probability of the outcome. A negative
regression coefficient means that predictor decreases the probability of that
outcome; a large regression coefficient means that the predictor strongly
influences the probability of that outcome; while a near-zero regression
coefficient means that that predictor has little influence on the probability
of that outcome. To each of these regressors a significance value is as-
signed that expresses the probability of rejecting the null hypothesis at a
particular significance level when the null hypothesis is actually true.
A caveat is in place, however. Logistic models rely on the assumption
that all observations are independent. This is not the case in a dataset where
three writers/translators provide all of the examples. In addition, not all
sources of repetition are of equal interest at this stage of the research. Given
that this study is the first quantitative study on the interaction between as-
pect and modality in Russian, focus will be on general tendencies – such as
finding out how a particular modality or polarity type interacts with the
aspect of the infinitive – rather than on the peculiarities of aspectual choice
after a particular modal adverbial predicative or in a specific novel. This
stance requires generalizing to the entire population of novels and modal
adverbial predicatives.
Variables such as novels and modal adverbial predicatives can be re-
garded as random effects because interest is in the underlying population,
not in the details of their behaviour. Here, focus is on quantifying the
amount of variability between adverbs, not in identifying individual effects;
individual differences are considered to be “random disturbances” (Searle,
Casella, and McCulloch, 1992: Section 1.4). Parameters that are interesting
in themselves, such as modality and polarity type, would qualify as fixed
effects. Bear in mind that the results pertain only to the specific levels of the
fixed effects distinguished (i.e. deontic versus dynamic, positive versus
negative) and cannot be extrapolated to any other theoretically plausible
levels.
One way of taking all these facts and factors into account is by fitting a
Mixed-Effects Logistic Regression model that allows us to control for fixed
and random individual differences (compare Bresnan et al. 2007). Follow-
ing Baayen (2008: chapter 7), I fitted a Mixed Effects Logistic Regression,
 Idiosyncratic aspect­modality relation in Russian 321

using the package ‘lmer’ in R 2.5.0, to the 314 Russian instances that con-
tain an infinitive existing in both imperfective and perfective, i.e. allowing
aspectual choice. The aim is to predict the propensity of modality type and
polarity type to co-occur with imperfective or perfective aspect. I will
summarize the results and draw theoretical conclusions in the fifth and final
section of this paper.

5.3. Results

The results of logistic mixed effects modelling show that in a model with
modality and polarity type as fixed effects, compared to the chance of en-
countering an imperfective infinitive, there is significantly less chance of
finding a perfective infinitive when deontic modality is expressed [coeffici-
ent estimate -5.4412, p=6.95e-11], yet significantly more chance of finding
a perfective infinitive when the modal (ad)verb is used in a positive sen-
tence [coefficient estimate 3.8689, p=0.000807]. In other words, both inde-
pendent variables, modality type and polarity type, play an important role
in unravelling the mystery of aspect assignment in modal constructions,
with modality type being the stronger predictor.
This outcome is not what was expected on the basis of the literature on
the aspect-modality interaction, however. Recall that, in the literature, mo-
dality type may well have been put forward as the best predictor for aspec-
tual choice, yet in doing so, imperfective aspect was linked to dynamic
modality. Three diagnostics confirm that we can nevertheless have trust in
this corpus-based model: the estimated scale of variation [0.9864484] is
very close to the ideal of [1]; in other words, there is only slightly less vari-
ation in the data than the model would expect. The C index of concordance
between predicted probabilities and observed responses equals
[0.8670398]. C is an accepted traditional performance metric for a Receiver
Operating Characteristic Curve (ROC), a standard technique for summariz-
ing classifier performance over a range of tradeoffs between true positive
(TP) and false positive (FP) error rates (Sweets 1988). When C takes on the
value 0.5 the predictions are random, and when it takes on the value 1 the
predictions are perfect; a value above 0.8 indicates that the model has real
predictive capacity, as it can be interpreted to mean that 80 percent of the
time a randomly selected individual from one group has a test value larger
than that for a randomly chosen individual from the other group. Despite
the fact that both true and false positives are taken into account when calcu-
322 Dagmar Divjak

lating C, the overall concordance rate of 86.7% is still considerably higher

than the 79.9% of utterances that would be classified correctly if no expla-
natory variables would be taken into account, and instead the perfective
would always be selected. Thirdly, Somer’s D, the rank correlation between
predicted probabilities and observed responses, ranges between 0 (random-
ness) and 1 (perfect prediction); in general a correlation is considered large
if it takes on a value over 0.5. The obtained [0.7340796] thus indicates a
large correlation between predicted probabilities and observed responses.
The outcome obtained suggests that the “lexical” meaning of modality
(dynamic vs deontic) as well as polarity (positive vs negative) predict as-
pectual choice in modal constructions quite well for Russian. Deontic mo-
dality and negative polarity favour the appearance of imperfective aspect,
whereas dynamic modality and positive polarity increase the chance of
encountering a perfective infinitive. Furthermore, both variables seem to
interact: it was found that, compared to the chance of finding an imperfec-
tive infinitive, chances of finding a perfective infinitive are significantly
lower when a negated deontic SoA is expressed.4

6. Discussion

Crucial for this paper is the fact that, thanks to a quantitative corpus-based
approach, objective and verifiable evidence was obtained for the existence
of a relation between a modal adverbial predicative followed by an imper-
fective infinitive and deontic modality. This finding reverses the claims
made in the typological linguistic literature while confirming the correc-
tions proposed by Russian linguists (see Section 2). Such a “reversed” out-
come is to be expected on the basis of what we know about the semantics
of the Russian aspectual system. For one, linguistic research on aspect in
Russian has always posited that, contrary to what is the case in many other
languages, in Russian, the imperfective is the unmarked aspect, not the
perfective (Forsyth 1970: 14). Recently proposed cognitive linguistic mod-
els of aspect allow to explain and predict the details of the present findings
quite naturally. Let us look at two such models.
Dickey (2000) uses concepts from Cognitive Grammar to construct a
semantic analysis of the category of aspect in the western group (Czech,
Slovak, Slovene, Sorbian), the eastern group (Russian, Ukrainian, Belaru-
sian, Bulgarian), and the transitional zone (Serbo-Croatian and Polish). On
Dickey’s account, the conceptual category underlying the Eastern aspectual
 Idiosyncratic aspect­modality relation in Russian 323

system is focused on definiteness in time, or lack thereof. Definiteness is

based on a pre-linguistic and cognitively basic perception of uniqueness,
which we experience constantly with respect to the cognitive domains of
both space and time (Dickey 2000: 266). In this system, the imperfective
expresses “qualitative temporal indefiniteness”, i.e. lack of assignability to
a single, unique point in time. This lack of uniqueness and assignability to a
single, unique point in time, so typical of imperfective aspect, fits well with
the “timeless applicability” that lies at the heart of deontic modality. The
more specific aspect, the one that is unique and assignable to a single point
in time, i.e. the perfective aspect, takes on the more restricted modal value,
dynamicity. It comes as no surprise then that the values of aspect and mo-
dality pair up the way they do in Russian, reversing tendencies that seem to
hold outside of the Slavic realm.
Janda (2004) approaches aspect in Russian from the point of view of
conceptual metaphor. Conceptual metaphors typically employ a more ab-
stract concept as target and a more concrete or physical concept as their
source. Building on this, Janda proposes that human experience of matter
provides the source domain for the metaphor motivating the grammatical
category of aspect in Russian. The model she proposes is a version of the
universal TIME IS SPACE metaphor, according to which SITUATIONS ARE
MATERIAL ENTITIES. Applied to aspect, this amounts to identifying PER-
FECTIVE with A DISCRETE SOLID OBJECT and IMPERFECTIVE with A FLUID
SUBSTANCE. The contrast of discrete solid objects with fluid substances
reveals over a dozen properties that can be divided into three groups: inher-
ent properties such as edges, shape, and integrity (which correspond to
inherent situation aspect); interactional properties such as juxtaposition,
dynamism, and salience (which correspond to discourse phenomena of
aspect); and human interactional properties such as graspability and impact
(which correspond to pragmatic phenomena of aspect). Bringing modality
into the equation reveals that the fluidity and shapelessness of imperfective
situations that fit in anywhere and fill up anything resembles the general
applicability typifying deontic modality quite well. The “limits” and
“boundaries” imposed on perfective situations, on the other hand, remind
more of the specific properties characterizing dynamic SoAs. Here too, the
way in which the values of aspect and modality line up reverses the general
tendencies, yet is motivated by and expected on the metaphorical model
that underlies Russian aspect.
Both cognitive linguistic models motivate the correlation between the
specific values of aspect and modality attracting each other quite naturally.5
324 Dagmar Divjak

Yet, which cognitive mechanism is responsible for this? The standard ap-
proach to the semantics of aspectual categories presents them as organized
around a prototype with extensions into other domains such as modality
(Boogaart and Janssen, 2007: 817). Rather than viewing modality as a
(metaphorical) extension from aspect, I propose considering the relation
between the two domains aspect and modality in terms of analogical map-
ping. In making analogies, we can follow at least three different cognitive
strategies that Holyoak and Thagard (1995) identify as 1) attribute mapping
or the perception or creation of similarity between objects, 2) relational
mapping or sensitivity to relations between objects, and 3) systems map-
ping or the recognition of patterns created by those object relations, which
enables generalization to more abstract structure. Freeman (2002: 468)
points out that, within Cognitive Linguistics, Fauconnier (1997) identifies
such mappings as projection, pragmatic function, and schema mappings.
Metaphors are examples of projection mappings, which project part of the
structure of one domain onto another. Pragmatic function mappings allow
an entity to be identified in terms of its counterpart in the projection and
thus give rise to metonymy and synecdoche. Schema mappings operate
when a general schema, frame, or model is used to structure a situation in
context and are the structural basis for allegory and symbolism.
As I mentioned above, it does not seem to be the case that modality is
understood metaphorically in terms of aspect, or vice versa. Situations that
are morally desirable are not easily thought of as lacking temporal boundar-
ies or being repeated. Vice versa, dynamic modality does not easily remind
of temporal boundedness. The similarities between the two domains are to
be found at a deeper level. Imperfective situations spread out over time,
while perfective situations are limited to a certain point in time. Corres-
pondingly, dynamic situations are restricted by participants or circum-
stances while deontic states of affairs apply regardless of any boundaries.
What aspect and modality have in common is the fact that both perfective
aspect and dynamic modality are bounded by someone or something and
hence marked in some respect, while imperfective aspect and deontic mo-
dality are unmarked and unbounded in every respect. Given that the proper-
ties of aspect and modality themselves are far from salient, seeing this simi-
larity requires drawing on knowledge domains of Idealized Cognitive
Cultural Models (Freeman 2002: 469), in particular on knowledge about the
way in which time affects situations and the written and unwritten code that
rules our lives. Mapping thus occurs at the systems or schema level as the
relations between the components of the mappings are more abstract and it
 Idiosyncratic aspect­modality relation in Russian 325

is the structure of the relations that is seen to be the same or isomorphic.

Furthermore, it is through schema abstraction that analogy can promote the
formation of abstract rules (Gentner and Medina 1998), which is precisely
what seems to have happened in the area of aspect assignment in modal
contexts. Faced with the need of having to make an aspectual choice in a
modal context, speakers of Russian were guided by the deep similarities
between the two domains. This, in turn, has given rise to the very strong
tendency to use the imperfective aspect to signal deonticity, and the perfec-
tive to render dynamic modality.

Notes

1. This paper is a written up version of the Russian data presented at the ICLC-
2007. I am grateful to the Science Foundation – Flanders (Belgium) for fund-
ing this project (2005-2008) and to the research unit QLVL for hosting it at
the KULeuven (Belgium); to Östen Dahl for countless discussions on aspect
and modality during and after my stay at Stockholms Universitet (2005-2006);
to Ruprecht von Waldenfels (University of Bern) for “sharing” the load of
automatically annotating, lemmatizing and aligning the corpus texts; to
Oesten Dahl, Leonid Oknjanskij and Nicholas LeBlanc for second-tagging
some of the Russian data; to Harald Baayen for providing me with a draft of
“Analyzing Linguistic Data”; to Kostas Triantafyllopoulos and Jean Russell
(University of Sheffield) for advice on (Mixed Effects) Logistic Regression
modeling; to the ICLC-2007 audience and the “general linguists” at Stock-
holms Universitet for valuable questions and comments.
2. Since the individual modal subtype seems to be irrelevant in predicting aspec-
tual choice, Padučeva (2006) proceeds to explaining the aspectual choices as
referring to the difference between a singular event in the future (triggering
perfective) versus multiple events in the present (requiring imperfective),
which corresponds partially with the generic/specific parameter that I will
introduce.
3. This assumption is confirmed by the fact that a model fit to data originally
produced in Russian alone, i.e. extracted from the novel Master i Margarita
by Bulgakov, yields exactly the same results as a model fit to a dataset that
consist of both original and translated utterances. This outcome was expected
by the fact that neither language, nor author/translator nor novel came up as
significant factors in the combined analysis. In other words, and contrary to
what many linguists might assume intuitively, whether or not original utter-
326 Dagmar Divjak

ances only are taken into account does not influence the results in any way.
Thanks to Kerstin Fischer for suggesting this course of action.
4. A second model, fit to assess how the model fares when modality type and
polarity type are combined, predicts that chances of finding a perfective in-
finitive are significantly lower when a negated deontic SoA is expressed. The
rule-like nature of the finding, i.e. the zero occurrences of this particular vari-
able constellation, makes statistical testing difficult. In order to circumvent
this problem, the dataset was restructured by merging the variables modality
type and polarity type into one, yielding the variable polarized modality type.
A mixed effects logistic regression model with novel and modal word as ran-
dom effects shows that chances of finding a perfective infinitive increase sig-
nificantly when the SoA is dynamic and positive, indeed [coefficient esti-
mate=3.7342, p=0.000866 in a model with an estimated scale of variation of
0.9716332, a C equal to 0.8680516 with 79.2% correct classifications if no
explanatory variables would be taken into account and Somer’s D being
0.7361032]. Thanks to Dr N. Fieller and Dr J. Russell (University of Shef-
field, Statistical Services Unit and CICS) for their advice on this issue
5. Leaving aside the question of whether modal meaning can and should be
considered as an extension of aspectual meaning (see Divjak 2009) and as-
suming a radial network of interrelated aspectual meanings, it could be argued
that the association between imperfective aspect and deontic modality is like-
wise expected on the traditional approach (Maslov 1959, Rassudova 1968).
This is less straightforward than motivating the correlation between a particu-
lar modality type with a particular aspect on the cognitive approach, however.
On the traditional view of aspect, deontic modality can be regarded as com-
patible with the imperfective’s central general-factual meaning, whereas dy-
namic modality can be said to relate to the perfective’s specific-factual mean-
ing. Deontic modality regulates existence for everyone, always and
everywhere, hence expresses a meaning that is similar to the prototypical in-
terpretation of the imperfective as encoding statements of fact, as events with
focus on the process or as repeated events. Dynamic modality is concerned
with a particular situation or a participant in that situation, hence quite similar
to the prototypical interpretation of perfectively coded events as having sum-
marizing properties and as presenting situations as one-off events or as events
with specific settings.
 Idiosyncratic aspect­modality relation in Russian 327

References

Abraham, Werner
1991 Modalverben in der Germania. In Begegnung mit dem ‘Fremden’:
Grenzen – Traditionen – Vergleiche. Akten des VIII. Internationalen
Germanisten-Kongresses in Tokio 1990, Vol. 4, E. Iwasaki (ed.),
109-118. München: Iudicum Verlag.
Abraham, Werner and Elisabeth Leiss
2006 Cross-linguistic aspect/Aktionsart-modality links. Workshop orga-
nized at the Conference of the Societas Linguistica Europaea, Sep-
tember 2006, Bremen.
Baayen, R. Harald
2008 Analyzing Linguistic Data: A practical introduction to statistics
using R. Cambridge: Cambridge University Press.
Bermel, Neil
1997 Context and the Lexicon in the Development of Russian Aspect,
(University of California Publications in Linguistics 129). Berkeley /
London: University of California Press.
Boogaart, Ronny and Theo Janssen
2007 Tense and aspect. In The Oxford Handbook of Cognitive Linguistics,
Dirk Geeraerts and Hubert Cuyckens (eds.), 803-828. Oxford / New
York: Oxford University Press.
Bresnan, Joan, Anna Cueni, Tatiana Nikitina and R. Harald Baayen
2007 Predicting the dative alternation. In Cognitive Foundations of Inter-
pretation, G. Bouma, I. Kraemer and J. Zwarts (eds.), 69-94.
Amsterdam: Royal Netherlands Academy of Arts and Sciences.
Bybee, Joan, Revere D. Perkins and William Paglucia
1994 The Evolution of Grammar: Tense, aspect, and modality in the lan-
guages of the world. Chicago: University of Chicago Press.
Buhusi, Catalin V. and Warren H. Meck
2005 What makes us tick? Functional and neural mechanisms of interval
timing. Nature Reviews Neuroscience 6: 755-765.
Comrie, Bernard
1976 Aspect: An introduction to the study of verbal aspect and related
problems. Cambridge: Cambridge University Press.
Dickey, Stephen
2000 Parameters of Slavic Aspect: A cognitive approach. (Dissertations in
Linguistics.) Stanford: Center for the Study of Language and Infor-
mation.
328 Dagmar Divjak

Divjak, Dagmar and Östen Dahl

2007 Quantifying the interaction between modality and aspect in Slavic
languages. Presented at 10th International Cognitive Linguistics
Conference, July 2007, Kraków.
Divjak, Dagmar
2008a Co motywuje wybór aspektu w konstrukcjach modalnych typu
można, wolno, trzeba, należy itp. Paper presented at
Językoznawstwo kognitywne 2008 (Cognitive Linguistics in the
Year 2008), April 2008, Łódź .
2008b Modeling aspectual choice in Polish modal constructions. A corpus-
based quest for the Holy Grail? Presented at Quantitative Investigat-
ions in Theoretical Linguistics 3, June 2008, Helsinki.
2009 Mapping between domains. The aspect-modality interaction in Rus-
sian. Russian Linguistics 33 (3): 249-269.
forthc. Aspect and modality in Slavic: An exceptional interaction.
Fauconnier, Gilles
1997 Mappings in Thought and Language. Cambridge: Cambridge Uni-
versity Press.
Forsyth, J.
1970 A Grammar of Aspect. Usage and meaning in the Russian verb.
Cambridge: Cambridge University Press.
Freeman, Margaret
2002 Cognitive mapping in literary analysis. Style 36(3): 466-83.
Gentner, Dedre and José Medina
1998 Similarity and the development of rules. Cognition 65: 263–297.
Goldbeter, Albert
2008 Biological rhythms: Clocks for all times. Current Biology 18(17):
R751-753.
Holyoak, Keith J. and Paul Thagard
1995 Mental Leaps: Analogy in creative thought. Cambridge, MA: MIT
Press.
Janda, Laura A.
2002 The conceptualization of events and their relationship to time in
Russian. Glossos 2 [http://www.seelrc.org/glossos].
2003 A user-friendly conceptualization of aspect. Slavic and East Euro-
pean Journal 47(2): 251-281.
2004 A metaphor in search of a source domain: the categories of Slavic
aspect. Cognitive Linguistics 15(4): 471-427.
2007 Aspectual clusters of Russian verbs. Studies in Language 31(3):
607-648.
 Idiosyncratic aspect­modality relation in Russian 329

Maslov, Jurij S.
1959 Glagol’nyj vid v sovremennom bolgarskom jazyke. Voprosy gram-
matiki bolgarskogo literaturnogo jazyka. Moskva: Izd-vo AN SSSR.
Nuyts, Jan, Pieter Byloo and Janneke Diepeveen
2005 On deontic modality, directivity, and mood. A case study of Dutch
mogen and moeten. Antwerp Papers in Linguistics 110.
[http://webhost.ua.ac.be/apil/apil110/apil110.pdf – last accessed on
15 February 2010].
Nuyts, Jan
2005 The modal confusion: On terminology and the concepts behind it. In
Modality: Studies in form and function, Alex Klinge and Hen-
rik Høeg Müller (eds.), 5-38. London: Equinox Publishing.
2006 Modality: Overview and linguistic issues. In The Expression of Mo-
dality, William Frawley (ed.), 1-26. Berlin: Mouton.
Padučeva, Elena V.
2006 Modality, negation and aspect: The case of the Russian možet and
dolžen. Conference of the Societas Linguistica Europaea, September
2006, Bremen.
Rassudova, Ol’ga P.
1968 Upotreblenie vidov glagola v russkom jazyke. Moskva: Izdatel’stvo
Moskovskogo Universiteta.
Searle, Shayle Robert, George Casella and Charles E. McCulloch
1992 Variance Components. New York: John Wiley and Sons.
Smith, Carlota
1997 The Parameter of Aspect, 2nd ed. Kluwer Academic Press.
Šmelev, Andrej and Anna A. Zaliznjak
2000 Vvedenie v russkuju aspektologiju. Moskva: Jazyki russkoj kul’tury.
2006 Aspect, modality and closely related categories in Russian. Paper
presented at the Inaugural Meeting of the Slavic Linguistic Society,
September 2006, Bloomington.
Trnavac, Radoslava
2006 Aspect and subjectivity in modal constructions. (LOT Dissertation
Series.) Leiden: LOT.
Wiemer, Björn
ms. Aspect choice in modal and pragmatic contexts: pieces of a puzzle in
Russian and other Slavic languages.
Wälchli, Bernhard
2007 Advantages and disadvantages of using parallel texts in typological
investigations. In Parallel Texts: Using translational equivalents in
linguistic typology (Special ed. Language Typology and Universals
60), Michael Cysouw and Bernhard Wälchli (eds.), 118-134. Biele-
feld: Akademie.
330 Dagmar Divjak

Corpus

ParaSol – Parallel Corpus of Slavic and Other Languages

<http://www.uni-regensburg.de/Fakultaeten/phil_Fak_IV/Slavistik/RPC/>
last accessed 15 February 2010.
Bulgakov, Michail
1938 Master i Margarita.
(Polish) Mistrz i Małgorzata (Irena Lewandowska and Witold Dąbrowski)
(Serbian) Мајстор и Маргарита (Milan Čopić)
Lem, Stanisław
1961 Solaris
(Russian) Солярис (Dmitrij Bruškin)
(Serbian) Solaris (Predrag Obućina)
Pavić, Milorad
1984 Hazarskij Rečnik.
(Polish) Słownik chazarski (Elżbieta Kwaśniewska, Danuta Cirlić-Straszyńska)
(Russian) Хазарский словарь (Larisa Savel’evaja)
Section IV

Towards an empirical Cognitive Semantics

Quantitative approaches in usage-based Cognitive
Semantics: Myths, erroneous assumptions, and a
proposal

Stefan Th. Gries and Dagmar S. Divjak

Abstract
In this paper, we assess objections formulated against (quantitative) corpus-
linguistic methods in Cognitive Linguistics. We present claims critical of both
corpus linguistics in general and particular corpus-linguistic analyses in particular
and discuss a variety of theoretical as well as empirical shortcomings of these
claims. In addition, we summarily discuss our recent corpus-based Behavioral
Profile approach to Cognitive Semantics and illustrate its advantages in the do-
mains of synonymy, polysemy, antonymy, and cross-linguistic semantics as well as
its methodological flexibility.
Keywords: Cognitive Semantics, corpus linguistics, polysemy, synonymy, behav-
ioral profile, English, Russian, statistical methods

1. Introduction

One of the most distinctive characteristics of Cognitive Linguistics is the

prominent role that meaning and function play in linguistic analyses. This
is true in two respects: first, in the sense that meaning and function are used
to explain phenomena traditionally regarded as belonging to the domain of
autonomous syntax; second, in the sense that studies of meaning and func-
tion themselves constitute a large body of cognitive linguistic work. At the
very beginning, the methodology underlying cognitive-linguistic studies
was quite homogeneous: just as in formal linguistics, analyses were nearly
exclusively based on the acceptability or appropriateness of utterances that
were often pulled out of their natural context and judged by the analyst
him/herself. This admittedly rather unfortunate state of affairs began to
change around the early 1990s when both experimental and observational
approaches became more frequent. The change mirrored to some degree a
general movement towards more rigorous empirical methods in linguistics,
but was also facilitated from within Cognitive Linguistics itself by the re-
334 Stefan Th. Gries and Dagmar S. Divjak

cognition that the hallmark of Cognitive Linguistics – extremely fine-

grained studies of the semantics of lexical items (in particular function
words such as the classic studies of over and there) – suffered from a vari-
ety of methodological shortcomings (cf. Sandra and Rice 1995 for discus-
sion). As a result, experimental research became more common in Cogni-
tive Linguistics, in particular in idiom research by Gibbs and colleagues. It
took another few years until observational approaches using corpus data
really took off, however. Their increasing popularity was no doubt facili-
tated by a rise of “usage-based approaches”. As a result frequency-based
accounts and corpus-linguistic methods gained ground, both in terms of
publications, theme sessions at Cognitive Linguistics conferences, and in
general visibility.
In spite of the apparent omnipresence of advocates of usage-based ap-
proaches, corpus-linguistic approaches are still much less widespread in
Cognitive Linguistics than one might be lead to believe. There are probably
several reasons for this fact:
– new methodologies have never caught on fast in linguistics;
– corpus data are not always available and, if available, they tend to be
difficult to handle, which is a friendly way of saying “introspective
judgments of made-up data are so much easier to generate, keep
track of, annotate, and evaluate” than thousands of diverse matches
extracted from a corpus;
– corpus-linguistic assumptions and methods are often poorly under-
stood;
– the quantitative underpinning of modern corpus-linguistic work does
not come natural to linguists as, since the middle of the last century,
the discipline has lacked a strong methodological foundation or even
awareness, unlike psychology or psycholinguistics.
These issues conspired to yield yet another, attitudinal, obstacle to a fast
rise of corpus-linguistic methods: the conviction that corpus-linguistic
methods have little to offer that the ‘good ol’ traditional approach’ could
not already do. We believe and will argue that this conviction, just like the
issues that gave rise to it, are entirely misguided. First, there is plenty of
evidence suggesting introspective judgments come with a variety of diffi-
culties that make them less objective, testable, and replicable than is desir-
able (cf. Labov 1972, 1975 as well as Schütze 1996 and the many studies
cited therein). Second, the absence of quantitative underpinning makes it
more difficult to compare the results obtained with those of other studies,
 Quantitative approaches in usage-based cognitive semantics 335

while some issues in linguistics are too complex and multidimensional in

nature to allow analysis by mere introspection (cf. Gries 2003, Hinrichs and
Szmrecsanyi 2007). Surely, psychologists and psycholinguists, who look at
issues just as complex and noisy as linguists, jump through methodological
and statistical hoops just for the fun of it.
In this largely programmatic and slightly polemic paper, we argue (i)
against recent and less recent yet recurrent criticisms of corpus-linguistic
methods and (ii) in favor of a particular corpus-based method for cognitive-
semantic analyses, the so-called Behavioral Profile approach (extending an
expression first used by Hanks 1996). More specifically, in the following
section we will first look at one attack against corpus-linguistic methods
that we consider representative albeit somewhat extreme. Next, we briefly
present our answer to this criticism, i.e. the Behavioral Profile method as
well as outline and exemplify several of the advantages it has over compet-
ing approaches. Finally, we also address a variety of arguments often
hurled at corpus-based approaches (within and outside Cognitive Linguist-
ics) to bolster our point that corpus-based methods are at the same time the
most useful and most underestimated methodological tool available to con-
temporary (cognitive) linguistics.

2. Discussion

2.1. Criticism targeted at corpus-linguistic methods in Cognitive

Linguistics

The most vociferous critic of corpus-based methods in Cognitive Linguist-

ics we have come across is Raukko (1999, 2003). In two papers on English
get, he argues vehemently against corpus-based methods in Cognitive Se-
mantics (and in favor of his own experimental method). While a full-
fledged rebuttal of the myriad of problems in his argumentation is not our
concern here (cf. Berez and Gries 2009 for that), his work exemplifies at
least some of the above-mentioned problems. One of these is Raukko’s
(mis)conception of corpus linguistics. This is how he characterizes the cor-
pus-linguistic method:
The linguist looks at a large and somewhat pre-processed selection of text
material and tries to find the relevant instances (instantiations, specimens)
of the item that s/he wants to study. (Raukko 2003:165)
336 Stefan Th. Gries and Dagmar S. Divjak

This statement is either a redundant truism, a severe misunderstanding,

or just as malign a misrepresentation. It is a redundant truism in the sense
that, sure, if a corpus linguist investigates get in a corpus, he only looks for
“relevant instances”, i.e. instances of the verb (lemma) get and not for the
noun formaldehyde. It is a severe misunderstanding or misrepresentation to
think that a corpus linguist worthy of the name would look for instances of
get in the corpus, yet would only classify those instances as relevant that do
fit his theory instead of classifying all instances found or at least a represen-
tative randomized sample, in order to avoid having to deal with problematic
instances and/or potential counterexamples. Put differently, contemporary
corpus linguistics does not restrict itself to selecting those examples that fit
a theory (an approach Tummers et al. (2005) refer to as corpus-illustrated
research), disregarding the rest – on the contrary, it is a strength of the cor-
pus-based approach, to which everybody who has ever looked at authentic
data can testify, that a comprehensive corpus search typically results in data
no introspection would have yielded and that all of these data are taken into
account.
Raukko (1999: 87) likewise takes issue with the fact that corpus lin-
guists use introspection in their analysis of corpus data:
Other types of recent analyses of lexical polysemy […] have made use of
language corpora as sources of real-life data, but here also the analyst basi-
cally relies on her/his own linguistic introspection when analyzing the in-
stances of a word in the texts and classifying them into neat semantic cate-
gories.
Again, this statement is either a redundant truism, a severe misunderstand-
ing, or a misrepresentation. Of course, the analysis of corpus data requires
classificatory decisions which are not always entirely objective – no corpus
linguist in his right mind would deny this fact, just as no scientist in the
humanities or social scientists would deny that some degree of intuition
plays a role in nearly any study. The real issue is that corpus data often
contain examples an armchair linguist would not think of and, thus, force
the researcher to take a broader range of facts into consideration. In addi-
tion, the concordance lines of a particular search expression and the uses of
a word and their frequencies constitute an objective database of the kind
that made-up sentences do not, since researchers cannot invent all the uses
of an expression in a corpus let alone their frequencies of occurrence. Thus,
even if the classification of the data points is not always maximally objec-
tive, at least their nature, scope, and amount is, and the ideas underlying the
annotation of examples can – and should – be made explicit. In addition,
 Quantitative approaches in usage-based cognitive semantics 337

the corpus linguist will strive to analyze the entire set of to some extent
subjectively annotated examples in an objective way (a point to which we
1
will return later), postponing intuition until the stage of interpretation.
Finally, there are linguists who argue that introspection should be the
central method of, say, Cognitive Semantics. In Talmy’s (2000: 4-5) words,
“[c]ognitive semantics is thus a branch of phenomenology […] the only
instrumentality that can access the phenomenological content and structure
of consciousness is that of introspection”, but results from introspection
“must be correlated with those resulting from other methodologies” such as
corpora, experimentation, and others. Again, we believe that subjective
judgments are inevitable to some extent, yet the questions arise of when
and how these judgments should be obtained and used. Talmy’s (2000: 6)
argument that introspection “is already a necessary component in most of
linguistics”, e.g., in syntactic grammaticality judgments, is beside the point.
The fact that many linguists have used introspection in the past does not
mean there are no problems associated with researchers providing both
theory and data, as we summarily discussed above. In both experimental
and corpus-based studies, the primary source of data is not the analyst him-
self; it is a truism that data must still be interpreted, yet as many steps as
necessary should be taken to avoid subjective biases, and theory-formation
needs to be kept at least one step away from the retrieval of the data. We
again ask: if nearly all cognitive psychologists and psycholinguists realize
this, why is this so hard for many a linguist? True, corpus linguistics studies
meaning in terms of use, which in turn is made tangible through distribu-
tion, and hence lends itself better to quantification. Corpus-based ap-
proaches to meaning may not be able to capture the essence of abstract
feelings like love or faith, but do other disciplines, typically considered to
be better geared for this task, fare better in this respect? Has philosophy or
religion come up with a generally accepted definition of either love or
faith? And how well do these disciplines describe and predict when and
how these concepts are used in everyday life? We strongly believe that
Cognitive Linguistics can only benefit from reducing the subjective ele-
ment in its methods as much as is feasible, and the methods and arguments
presented below attempt to take important steps in this direction.
338 Stefan Th. Gries and Dagmar S. Divjak

2.2. The Behavioral Profile (BP) approach

2.2.1. Introduction

As a corpus-based approach, the BP approach is based on the truism that

corpus data provide (nothing but) distributional frequencies. A more rel-
evant assumption, however, is that distributional similarity reflects, or is
indicative of, functional similarity; our understanding of functional simi-
larity is rather broad, i.e., encompassing any function of a particular expres-
sion, ranging from syntactic over semantic to discourse-pragmatic. The BP
method involves the following four steps:
– the retrieval of (a representative random sample of) all instances of a
word’s lemma from a corpus in their context (usually at least the
complete utterance/sentence);
– a (so far largely) semi-manual analysis of many properties of the use
of the word forms; these properties are, following Atkins (1987), re-
ferred to as ID tags and comprise
– morphological characteristics of the usage of the word in
question: tense, aspect, mood, voice, number marking, etc.;
– syntactic characteristics of the usage of the word in question:
use in main or subordinate clauses, sentence type;
– semantic characteristics: the sense of the word, semantic
roles of the word’s arguments and adjuncts;
– the generation of a co-occurrence table that specifies which ID tag
level is attested how often with each word (of a set of near synonyms
or antonyms) or sense (of a polysemous word) as well as the conver-
sion of these observed frequencies into within ID-tag percentages.
– the evaluation of the table by means of descriptive techniques (such
as summary frequencies), correlational methods, and exploratory
cluster analysis.
To clarify, what we refer to as a Behavioral Profile of a word or a word’s
sense is the column containing the percentages of (co-)occurrence for that
word or sense; consider the percentage columns in Table 1 for examples.
That also means that a Behavioral Profile is a data-driven operationaliza-
tion of a lexeme (or a word form)’s behavior in a corpus: we assume, as
mentioned above, that semantic, pragmatic, and other kinds of differences
between words or senses will be reflected in different distributions of for-
 Quantitative approaches in usage-based cognitive semantics 339

mal and other linguistic features, so that the BP is of diagnostic for func-
tional differences.
Table 6. An excerpt of the behavioral profiles for three ID tags of begin and start

ID tag begin start

name levels n % n %
declarative 290 0.9732 511 0.9623
sentence type interrogative 6 0.0201 12 0.0226

imperative 2 0.0067 8 0.0151

main 135 0.453 231 0.435
clause type
dependent 163 0.547 300 0.565

semi 128 0.4295 91 0.1714

copula 0 0 1 0.0019

transitive 0 0 2 0.0038
verb type
monotransitive 34 0.1141 92 0.1733

intransitive 118 0.396 243 0.4576

semip 18 0.0604 102 0.1921

In the following section, we discuss and exemplify several applications.

The examples involve all kinds of statistical methods as well as examples
from the domains of polysemy and near synonymy, within one language as
well as across languages.

2.2.2. Applications in polysemy

Gries (2006a) uses descriptive methods based on behavioral profiles of the

senses of run to address several of the central questions Cognitive Semanti-
cists face. For example, Gries (2006a: Section 4.1) addresses the question
of identifying the prototypical sense of run on the basis of several criteria
including the most frequent sense and the formally least marked or con-
strained sense. Obviously, the BP approach allows operationalizing these
two criteria straightforwardly. Classifying all concordance lines per verb
340 Stefan Th. Gries and Dagmar S. Divjak

sense makes it possible to count which sense is the most frequent one; the
formally least constrained sense can be defined as the sense that is encount-
ered with the largest attested number of ID tag levels (corrected for sense-
frequency). Both criteria point to the sense ‘fast pedestrian motion’, which
is not only intuitively correct, but also supported by other corpus-based,
though not BP-based, findings such as the fact that this sense is both onto-
genetically privileged (i.e., acquired first by children), phylogenetically
privileged (i.e., one of two diachronically earliest senses) and, in addition,
the most frequent sense of the zero-derived noun run.
BPs also answer the question of where to connect a particular sense of a
polysemous word to the network of already identified senses forming a
network. The example in question deals with the senses ‘move away from
something dangerous/unpleasant’ and ‘move away to engage in a romantic
2
relationship’; the three most likely – ‘most likely’ in the sense that they are
semantically most similar – points of connection are the senses ‘fast pedes-
trian motion’, ‘fast motion’, and ‘motion’. All other things being equal,
Gries suggests to base one’s decision of which two senses to connect on the
overall distributional similarities between the two senses and between the
candidate senses recognized in the network. The overall distributional simi-
larity between two senses is operationalized as the correlation coefficient of
the two senses’ behavioral profiles. An investigation of all correlations
between all senses shows that
– the five senses of run in question are much more similar to each
other than all senses are to each other on average;
– the two senses that need to be connected are significantly more
similar to ‘fast pedestrian motion’ than to ‘fast motion’ and ‘mo-
tion’, so this is how the network structure should be devised (again,
in the absence of additional evidence to the contrary).
An example of a cluster-analytic approach in the domain of polysemy is
Berez and Gries (2009). They investigate the senses of the highly polysem-
ous verb get in a small sample of the the ICE-GB using the BP approach.
They run a hierarchical agglomerative cluster analysis on the data and cal-
culate p-values based on multiscale bootstrap resampling (cf. Shimodaira
2004, Suzuki and Shimodaira 2006). In spite of the small sample size, they
find
– a cluster with all ‘possess’ senses (p≈0.07 marginally significant);
– a cluster with all the ‘acquire’ senses (p≈0.1 marginally signifi-
cant);
 Quantitative approaches in usage-based cognitive semantics 341

– all non-causative ‘move’ senses(p≈0.03 significant);

– a cluster that contains all causative senses (but also two other
senses; p≈0.21 not significant);
– a cluster that contains both grammaticalized senses ‘must’ and the
get-passive (but also one other sense; p≈0.08 marginally signifi-
cant).
Four out of five clusters are at least marginally significant, which is a good
result given a small sample size and the fact that clustering is after all an
exploratory method. The results therefore provide support for the fact that
distributional characteristics are strongly correlated with semantic charac-
teristics and senses of words, which in turn is exactly the assumption on
which the BP approach is based.

2.2.3. Applications in near synonymy

In the domain of near synonymy, Divjak (2006) investigates five verbs that
express ‘intend’ in Russian, whereas Divjak and Gries (2006) investigate
nine Russian verbs meaning ‘try’ on the basis of the verbs’ behavioral pro-
files. More specifically, the first study uses the BP approach to address the
delineation (which verbs should be considered near-synonyms?) and struc-
turing problem (how should a set of near synonymous words be struc-
tured?), whereas the second study focuses on the structuring and descrip-
tion problem (how can different words' meanings be compared reliably?).
Let us take a brief look at the try-study. On the basis of nearly 1,600
concordance lines of the verbs, their cluster-analytic approach reveals a
tripartite cluster structure, as well as several interesting differences between
the three clusters that are hard to discern in any other way. On the basis of
t-values reflecting between-cluster differences, it is shown, for example,
that the ID tag levels of each cluster give rise to a different abstract scen-
ario (cf. Divjak and Gries 2006: 42ff. and esp. Divjak 2010, for details):
– the cluster {pytat’sja, starat’sja, probovat’}: a human being is ex-
horted to undertaken an attempt to move himself or others (often
negated);
– the cluster {silit’sja, proyvat’sja, norovit’}: an inanimate subject
undertakes several repeated but non-intense attempts to exercise
physical motion;
342 Stefan Th. Gries and Dagmar S. Divjak

– the cluster {pyzit’sja, tuzit’sja, tscit’sja}: an inanimate subject at-

tempts in vain but intensely to perform what typically are meta-
phorical extensions of physical actions.
In addition, Divjak and Gries (2006) use z-values to identify within-cluster
differences because while, say, {silit’sja, proyvat’sja, norovit’} is a cluster
distinct from the other verbs, that does of course by no means imply that
the three are used identically.
An experimental follow-up study (Divjak and Gries 2008; for details,
see below) revealed that native speakers of Russian sort the nine verbs,
with the exception of silit’sja, into exactly the same clusters as the corpus-
based cluster analysis suggested. The fact alone that the BP approach is
able to structure the near synonyms in a way that is so similar to the ex-
perimental results and that no intuitive lexicographic analysis has suggested
so far shows how powerful corpus-based methods are and how useful the
BP approach is in particular for lexical and Cognitive Semantics.
Apart from these home-grown analyses that are BP analyses of near
synonymy per se, there are some studies which differ quantitatively from
the BP approach, yet not as much qualitatively. For example, Schmid’s
(1990) study of the phasal verbs begin and start involves a variety of ID
tags similar to the ones mentioned above; the main difference to the BP
approach is the degree to which statistical methods are used in the analysis.
Arppe’s (2007) and Arppe and Järvikivi’s (2007) studies of Finnish verbs
meaning ‘think’ is even more similar in that they involve both a very simi-
lar range of ID tags and multifactorial evaluation, but the foci are slightly
different: in both papers, the focus is exclusively on the differences be-
tween words, i.e., a more coarse-grained approach than investigating both
between-cluster and within-cluster differences, while the second paper is
more concerned with showing how experimental data supplement corpus
data (cf. below). Janda and Solovyev (2009) as well as other work by Janda
and colleagues employs a subset of the BP approach – basically some syn-
tactic ID tags that capture in which constructions a verb is used – to study
synonymous expression in Russian. Finally, Dąbrowska (2009) studies how
nine English verbs of bipedal motion group together on the basis of their
collocational patterns and semantic preferences. While she does not per-
form a detailed statistical analysis of the corpus results, our own cluster
analysis of her data (with the same settings we used for all BP cluster ana-
lyses to date) is to some extent compatible with her grouping.
 Quantitative approaches in usage-based cognitive semantics 343

2.2.4. Applications in antonymy and synonymy

In one recent study, Gries and Otani (forthc.) studied both antonymy and
synonymy in one and the same BP analysis using corpus data on the differ-
ent inflectional forms of the SIZE adjective lemmas big, great, large and
little, small, and tiny. They find that the BP approach suggests the canoni-
cally antonymous adjective pairs suggested in previous work (big vs. little
and large vs. small) for both a large number of semantic and a small num-
ber of syntactic ID tags. On amore theoretical level, they also argue that
Behavioral Profiles (i) are compatible with Hoey’s theory of lexical prim-
ing and a psycholinguistic exemplar-based approach to linguistic represen-
tation and processing and (ii) unify the distinction of the substitutability and
the co-occurrence approach in antonymy research.

2.2.5. Cross-linguistic studies

The studies mentioned in the previous section are all based on data from
one language. Cross-linguistic semantic studies are notoriously challenging
given that different languages carve up conceptual space(s) in different
ways (cf. Janda 2009 for discussion); for that reason linguistic dimensions
are difficult to compare across languages. Since the BP approach is based
on clearly operationalizable distributional properties, concordance lines
from different languages can be annotated for a number of common charac-
teristics while at the same time doing justice to any individual language’s
characteristics and avoiding overly subjective intuitions regarding cross-
linguistic semantic differences.
Divjak and Gries (2009) study near synonymous phasal verbs in English
(begin and start) and Russian (načat’, načat'sja, and stat’). As in other
applications, they annotate concordance lines for these five verbs for a
variety of criteria: morphological (tense, aspect, mode, person, voice), syn-
tactic (clause, sentence, and complement types), argument-structural prop-
erties, semantic roles of subjects and complements as well as verb sense.
Divjak and Gries then investigate these near synonyms within and be-
tween languages by comparing pairwise differences between the behavioral
profiles. For the within-language comparisons they find that the difference
between begin and start revolves around the semantic roles and character-
istics of the Beginner and the Beginnee, but that the main differences be-
tween the Russian verbs are not primarily concerned with the Beginner and
344 Stefan Th. Gries and Dagmar S. Divjak

the Beginnee. Rather, the verbs differ most strongly along aspectual and
argument-structural lines. Thus, for the between-language differences, Eng-
lish and Russian phasal verbs opt for a different division of the conceptual
space in question. While such cross-linguistic distinctions may be over-
looked in intuitive studies, they readily fall out from their behavioral pro-
files.
Just as in the previous section, conceptually similar work is available,
and Schönefeld (2006) is a case in point. She investigates translational
equivalents of three basic posture verbs in English, German, and Russian.
The main difference between her study and the BP approach outlined above
is that she includes only collocations to the exclusion of morphosyntactic or
semantic-role information. Another comparable study is Xiao and McEnery
(2006), who explore near synonyms from three lexical fields (the conse-
quence group, the cause group, and the price/cost group) on the basis of
their collocational behavior in English and Mandarin Chinese. Finally, De-
shors (forthc.) studies the different uses of the modals may and can in na-
tive-speaker English, English written by French learners, and the use of
pouvoir in French.

2.2.6. Further methods, validation, and converging evidence

Since behavioral profiles are based on distributional properties captured by

percentages, they offer possibilities that intuitive analyses lack: A final
attractive feature of the BP approach, therefore, is the fact that it allows
researchers to analyze the BP data using statistical techniques as well as to
compare the results to data/results from other studies. Armchair data is
much more limited in this respect.
For example, as we mentioned above, the BP approach was initially de-
veloped using descriptive, correlational, and exploratory, cluster-analytic
quantitative methods. However, given the variety of studies we have men-
tioned, it is obvious that different techniques may well be applied to the
type of data collected in a BP. To name just two examples: To find out
which variables drive the clustering Divjak (2010) uses a Linear Discrimi-
nant Analysis and in order to test the predictive power of the data contained
in the behavioral profiles for categorizing the verbs into clusters she fits a
logistic regression model. Arppe (2007) and Arppe and Järvikivi (2007)
also use logistic regression to predict the choice of one near-synonym over
another. To determine distinctive collocates, Schönefeld (2006) investigates
 Quantitative approaches in usage-based cognitive semantics 345

her data using hierarchical Configural Frequency Analysis (cf. von Eye
1990 and Gries 2004).
In addition, the quantitative nature of behavioral profiles allows for de-
tailed comparisons of BP based results with experimental evidence. For
example, Dąbrowska (2009) is concerned with how data from a forced-
choice selection tasks (of definitions and of video clips) and a gap-filling
task relate to the collocational data discussed above. Arppe and Järvikivi
(2007) discuss their corpus data with results from a forced-choice selection
task and an acceptability rating task.
More from the validation perspective, Divjak and Gries (2008) use a
gap-filling task and a sentence-sorting task to test their BP-based cluster
solution of the nine Russian verbs meaning ‘try’. In the gap-filling task,
subjects were given sentences from which the verb meaning ‘try’ had been
deleted and which exhibited ID tag levels strongly associated with one
verb. They then were asked to supply the verb they thought was most ap-
propriate for the sentence. In the sentence sorting task, subjects were given
sentences which differed only with regard to the try-verb used and were
asked to sort them into groups. Using Chi-square tests and a similarity met-
ric based on a Monte Carlo simulation, Divjak and Gries (2008) found that
the experimental findings are significantly more similar to the BP-based
cluster dendrogram than would be expected by chance, which lends strong
support to the assumption that the BP approach yields cognitively realistic
analyses.
By way of an interim summary, we have discussed applications and ad-
vantages of, as well as empirical evidence in favor of, the BP approach.
BPs can be used to investigate semantic relations of polysemy and syn-
onymy at a high level of granularity and objectivity; they can be applied to
simple cases with just two synonyms or larger sets with (so far) up to nine
synonyms, where analysts’ intuitions would become increasingly subjec-
tive, imprecise, and overtaxed; BPs allow to perform otherwise notoriously
difficult cross-linguistic studies, and given their quantitative nature, they
can be straightforwardly related to other empirical data and easily validated
experimentally. Whichever limitations there may still be, we believe that
the BP approach has much more to offer than many if not most other cur-
rently available approaches to lexical semantics, and certainly more than
some misguided and generic criticism of corpus-linguistic methods sug-
gests.
346 Stefan Th. Gries and Dagmar S. Divjak

2.2.7. Criticism targeted at specific aspects of corpus-linguistic methods in

Cognitive Linguistics

So far, we have mainly been concerned with presenting advantages of the

corpus-based BP approach in Cognitive Semantics (Sections 2.2.2 to 2.2.5)
and disarming general points of critique raised against using corpora in
Cognitive Semantics (Section 2.1). However, there is another set of argu-
ments, leveled at individual corpus-based studies, both outside of and
within Cognitive Linguistics. These can be summarized in what probably
are the two most frequently heard and most disliked remarks after corpus-
linguistic presentations:
– comments aimed at the corpus as a whole: “but isn’t all this true in
your corpus only?” or “you would most certainly find something
entirely different if you looked at a different corpus!” and “but the
two corpora you are comparing are not sufficiently similar, your re-
sults are invalid!”;
– comments aimed at a subpart of the corpus: “I bet you would find
something different if you looked at different registers!” or “I’m
sure you would find something different if you looked at word
forms/lemmas instead of lemmas/word forms.”
In spite of their frequency, these comments tend to be invalid. First, they
are theoretically problematic: The ‘asker’ hypothesizes a deviation from the
null hypothesis (that there is no effect of, or distributional difference be-
tween, corpora), i.e., an alternative hypothesis, yet places the burden of
proof on the ‘askee’. If the asker thinks the distributional data obtained and
reported on would be different in another corpus, the asker should test this
alternative hypothesis instead of stipulating a difference for which (so far)
no evidence exists; this is of course especially true when corpora on lan-
guages other than English are involved where alternative corpora are far
from easily available (if at all).
Second, assertions like these are empirically problematic: The kinds of
differences often hypothesized by askers is usually far from ‘a given’. As a
matter of fact, there now is an increasing amount of evidence that simple
generalizations of what does and what does not remain constant across
corpora, registers, word forms etc. are often inaccurate or exaggerated.
Some of this evidence is based on BP type of approaches, while other evi-
dence is based on data regarding the distribution of occurrences of syntactic
variables or the distribution of co-occurrences of lexico-syntactic variables.
 Quantitative approaches in usage-based cognitive semantics 347

As for the comments aiming at the corpus as a whole, for example, the
results obtained by Schmid (1993), who worked with the LOB corpus, are –
while less comprehensive in terms of annotation and more comprehensive
in terms of sense differentiation – to a considerably degree compatible with
Divjak and Gries’ (2009) results. This is noteworthy because the compo-
sition of the two corpora are of such a different nature that they would
compel many an audience to doubt the corpus comparability: Schmid’s
(1993) LOB consists exclusively of written and published texts representa-
tive for British English of the 1960s, whereas approximately 60% of the
ICE-GB corpus used in Divjak and Gries (forthc. a) consists of spoken
language and even the 40% of written language in the ICE-GB contains a
sizable amount of unpublished material.
Similar findings have been reported for the cherished distinction be-
tween spoken and written data. Stefanowitsch and Gries (2008) and Gries
(forthc.) show that distinguishing between spoken and written data has no
substantial effect in analyses of lexico-syntactic preferences of active vs.
passive voice, the two word orders of verb-particle constructions, and the
will vs. going-to future. Gries (forthc.) shows that the same holds true for
the ditransitive vs. prepositional dative alternation and that the ‘real’ divi-
sion of the ICE-GB corpus – ‘real’ in the sense of explaining the maximally
meaningful amount of variance in the corpus data as obtained by a Princi-
pal Component Analysis – cuts across both spoken vs. written and all regis-
ter distinctions present in the corpus. More specifically, the four corpus
parts that are most homogeneous internally and most different from each
other are based neither only on spoken vs. written nor only on subregisters;
instead, they are mixed groups based on both these levels of granularity.
This is of course something that linguists in general and linguists who have
never adopted a bottom-up approach to corpus data in particular would be
very reluctant to suggest; as scientists, they often prefer to stick to one level
of categorization … $100 to the first corpus linguist who gets the following
comment after a presentation: “Maybe you should forget about the mode
and the registers – I bet the real distinctions are actually a mixture of differ-
ent levels of corpus granularity.” Gries (forthc.) also finds that looking at
word forms does not necessarily yield results different from a lemma-based
analysis.
More generally, Gries (2006b) demonstrates on the basis of three very
different case studies – the frequencies of the present perfect, the predicta-
bility of particle placement, and lexicosyntactic associations of the ditransi-
tive constructions – that the usual suspects of mode, register and even sub-
348 Stefan Th. Gries and Dagmar S. Divjak

register account for much less variability than the above-mentioned after-
presentation comments suggest. In each of the above cases, different sam-
ples from even a single corpus may yield very different results; the size of
within-corpus differences is often similar in size to between-corpus differ-
ences so there is little reason to assume a priori that other corpora will
automatically yield different results. Bottom line: the issue of corpus
homogeneity and comparability can only be determined (i) empirically and
(ii) individually for each phenomenon, each corpus, and each level of cor-
pus division(s) – it cannot be determined or objected a priori as one sees fit
(cf. Gries, Hampe and Schönefeld 2010, for an example in Cognitive Lin-
guistics).

3. Concluding remarks

It goes without saying that this paper argues in favor of (in decreasing order
of generality) in linguistics in general and Cognitive Linguistics in particu-
lar:
– multi-methodological approaches;
– corpus-linguistic approaches;
– BP approaches.
While usage-based is one of the buzzwords in contemporary Cognitive
Linguistics, we believe that prototypical usage-based methods such as cor-
pus-linguistic methods are still underutilized, misunderstood, misrepre-
sented, and overcriticized, which is particularly interesting given that most
analyses based on subjective and unfalsifiable intuitions by a native
speaker-linguist are hardly ever subjected to any methodological critique …
As far as multi-methodological approaches are concerned, it is probably
fair to say that in linguistics as a whole the proportion of scholars combin-
ing different methods is increasing at a steady pace (cf. Gilquin and Gries
2009); fortunately, this development has carried over to Cognitive Linguist-
ics to some degree. The necessary development towards more quantitative
methods, however, is progressing at a slower pace. The need of statistical
methodological tools that are standard in most other social and cognitive
sciences (!) has not yet been recognized uniformly:
That is to say, a usage-based linguistics needs quantification and statistical
analysis. (Tummers, Heylen, and Geeraerts 2005:234)
 Quantitative approaches in usage-based cognitive semantics 349

The statistical analysis of empirical data, to be sure, should not be con-

sidered a fancy gadget designed to overwhelm linguists who are generally
not really acquainted with statistical techniques. Instead, statistical tech-
niques constitute an essential part of an empirical analysis based on corpus
data. (Tummers, Heylen, and Geeraerts 2005:236)
For corpus-linguistic approaches, we have pointed out some points of cri-
tique fairly commonly leveled against both corpus-linguistic methods in
general and corpus-linguistic studies in particular. We have addressed
every single of these points theoretically and/or on the basis of empirical
data and hope to have shown that these criticisms are often, though not
always, false, biased, not substantiated, and premature. In addition, a vari-
ety of advantages of corpus-based approaches has hopefully become appar-
ent: good corpus studies take into consideration all the variability that
comes with many natural examples (as opposed to few judgments on poten-
tially atypical examples), are gathered in an objective way and allow for
replicability and validation (which intuitive judgments do not).
As for the BP approach, we have argued that this radically corpus-based
approach, when applied to different kinds of semantic relations,
– yields more objective and more precise descriptive data than intro-
spective analyses while at the same time staying true to the usage-
based commitment of Cognitive Linguistics;
– allows for a bottom-up, data-driven study of distributional patterns
on the basis of many quantitative techniques that outperform hu-
man analysts in terms of pattern recognition;
– allows for cross-linguistic comparisons of lexical semantics using
objectively measurable distributional properties as opposed to diffi-
cult-to-port-across-languages semantic distinctions alone;
– allows integrating data and results from different sources and stud-
ies more easily than most other approaches (let alone intuitive ap-
proaches).
In a – we think – very positive sense, the BP approach is therefore usage-
based Cognitive Semantics in its most usage-based form. We invite the
skeptic who thinks that the good ol’ traditional way of doing semantic ana-
lyses can do all this and even more to illustrate how that is supposed to
work…
A final advantage of corpus-based approaches is that they are humbling;
humbling in the positive sense that the sober reality of what is attested in
corpora often puts a serious limit on bold theorizing. Put differently, the
350 Stefan Th. Gries and Dagmar S. Divjak

large degree of diversity corpus data exhibit as well as an indication of

what is frequent and what is not, of what does and what does not reach or
come close to reaching statistical significance, do not always support far-
reaching theoretical models and force practitioners to take the usage-based
perspective more seriously. Thus, given all the above, we plead for not
throwing out the corpus-linguistic baby with the argumentatively and
methodologically muddy bathwater and hope that the incredibly powerful
and flexible tool of quantitative corpus linguistics will become recognized
for what it can and what it cannot do.

Notes

1. Again, the fact that, in spite of Raukko’s critique of corpus-based methods, his
own approach fares no better in terms of objectivity will not be discussed
here; cf. again Berez and Gries (2009).
2. Cf. Gries (2006: section 4.2) for why these are considered different senses in
the first place.

References

Arppe, Antti
2007 Multivariate methods in the corpus-based lexicography: a study of
synonymy in Finnish. Paper presented at Corpus Linguistics 2007,
University of Central England, Birmingham.
Arppe, Antti and Juhani Järvikivi
2007 Every method counts: combining corpus-based and experimental
evidence in the study of synonymy. Corpus Linguistics and Lin-
guistic Theory 3(2): 131-59.
Atkins, Beryl T. Sue
1987 Semantic ID tags: Corpus evidence for dictionary senses. In Pro-
ceedings of the Third Annual Conference of the UW Centre for the
New Oxford English Dictionary, 17–36.
Berez, Andrea L. and Stefan Th. Gries
2009 In defense of corpus-based methods: A behavioral profile analysis of
polysemous get in English. In Proceedings of the 2008 Northwest
 Quantitative approaches in usage-based cognitive semantics 351

Linguistics Conference, 157–166. Seattle, WA: Department of Lin-

guistics.
Dąbrowska, Ewa
2009 Words as constructions. In New Directions in Cognitive Linguistics,
Vyvyan Evans and Stéphanie Pourcel (eds.), 201–223. Amsterdam /
Philadelphia: John Benjamins.
Deshors, Sandra C.
forthc. Interlanguage modality: A corpus-based study of may and can used
by French English learners. Unpublished Ph. D. dissertation, Uni-
versity of Sussex.
Divjak, Dagmar S.
2006 Ways of intending: A corpus-based cognitive linguistic approach to
near synonyms in Russian. In Corpora in Cognitive Linguistics.
Corpus-based approaches to syntax and lexis, Th. Gries and Anatol
Stefanowitsch, (eds.), 19–56. Berlin / New York: Mouton de
Gruyter.
2010 Structuring the Lexicon: A clustered model for near-synonymy.
Berlin / New York: Mouton de Gruyter.
Divjak, Dagmar S. and Stefan Th. Gries
2006 Ways of trying in Russian: Clustering behavioral profiles. Corpus
Linguistics and Linguistic Theory 2: 23–60.
2008. Clusters in the mind? Converging evidence from near-synonymy in
Russian. The Mental Lexicon 3: 59–81.
2009 Corpus-based Cognitive Semantics: A contrastive study of phasal
verbs in English and Russian. In Cognitive Corpus Linguistics Stud-
ies, Barbara Lewandowska-Tomaszczyk and Katarzyna Dziwirek
(eds.), 273–296. Frankfurt/Main: Peter Lang.
von Eye, Alexander
1990 Introduction to Configural Frequency Analysis: The search for types
and antitypes in cross-classifications. Cambridge: Cambridge Uni-
versity Press.
Gilquin, Gaëtanelle and Stefan Th. Gries
2009 Corpora and experimental methods: A state-of-the-art review. Cor-
pus Linguistics and Linguistic Theory 5: 1–26.
Gries, Stefan Th.
2003 Multifactorial Analysis in Corpus Linguistics: A study of particle
placement. London / New York: Continuum Press.
2004 HCFA 3.2. A program for R.
2006a Corpus-based methods and Cognitive Semantics: The many senses
of to run. In Corpora in Cognitive Linguistics. Corpus-based ap-
proaches to syntax and lexis, Stefan Th. Gries and Anatol Ste-
fanowitsch (eds.), 57–99. Berlin / New York: Mouton de Gruyter.
352 Stefan Th. Gries and Dagmar S. Divjak

2006b Exploring variability within and between corpora: some method-

ological considerations. Corpora 1: 109–51.
forthc. Corpus data in usage-based linguistics: What’s the right degree of
granularity for the analysis of argument structure constructions?. In
Expanding Cognitive Linguistic Horizons, Mario Brdar, Milena Žic
Fuchs and Stefan Th. Gries (eds.). Amsterdam / Philadelphia: John
Benjamins.
Gries, Stefan Th. and Naoki Otani
forthc. Behavioral profiles: A corpus-based perspective on synonymy and
antonymy. ICAME Journal.
Gries, Stefan Th. and Dagmar S. Divjak
2009 Behavioral profiles: A corpus-based approach to Cognitive Semantic
analysis. In New Directions in Cognitive Linguistics, Vyvyan Evans
and Stéphanie Pourcel (eds.), 57–75. Amsterdam / Philadelphia:
John Benjamins.
Gries, Stefan Th., Beate Hampe and Doris Schönefeld
2010 Converging evidence II: More on the association of verbs and con-
structions. In Empirical and Experimental Methods in Cogni-
tive/Functional Research, Sally Rice and John Newman (eds.), 59-
72. Stanford: CSLI.
Hanks, Patrick
1996 Contextual dependency and lexical sets. International Journal of
Corpus Linguistics 1: 75–98.
Hinrichs, Lars and Benedikt Szmrecsanyi
2007 Recent changes in the function and frequency of Standard English
genitive constructions: A multivariate analysis of tagged corpora.
English Language and Linguistics 11: 437–474.
Janda, Laura A.
2009 What is the role of semantic maps in Cognitive Linguistics? In Fest-
schrift for Barbara Lewandowska-Tomaszczyk, Piotr Stalmaszczyk
and Wieslaw Oleksy (eds.), 105–124. Frankfurt/Main: Peter Lang.
Janda, Laura A. and Valery D. Solovyev
2009 What constructional profiles reveal about synonymy: A case study
of Russian words for SADNESS and HAPPINESS. Cognitive Linguistics
20: 367–393.
Labov, William
1972 Some principles of linguistic methodology. Language in Society 1:
97–120.
1975 Empirical foundations of linguistic theory. In The Scope of Ameri-
can Linguistics, Robert Austerlitz (ed.), 77–133. Lisse: Peter de
Ridder.
 Quantitative approaches in usage-based cognitive semantics 353

R Development Core Team.

2007 R: A language and environment for statistical computing. R Founda-
tion for Statistical. Computing, Vienna, Austria. ISBN 3-900051-07-
0, URL <http://www.R-project.org>
Schütze, Carson T.
1996 The Empirical Base of Linguistics: Grammaticality judgments and
linguistic methodology. Chicago: University of Chicago Press.
Shimodaira, Hidetoshi
2004 Approximately unbiased tests of regions using multistep-multiscale
bootstrap resampling. Annals of Statistics 32: 2616–2641.
Stefanowitsch, Anatol and Stefan Th. Gries
2008 Register and constructional semantics: A collostructional case study.
In Cognitive Sociolinguistics, Gitte Kristiansen and René Dirven
(eds.), 129–152. Berlin / New York: Mouton de Gruyter.
Suzuki, Ryota and Hidetoshi Shimodaira
2006 The pvclust package, version 1.2.0, URL <www.is.titech.ac.jp
/~shimo/prog/pvclust/>.
Talmy, Leonard
2000 Toward a Cognitive Semantics. Vol. 1. Concept structuring systems.
Vol. 2. Typology and process in concept structuring. Cambridge,
MA: MIT Press.
2007 Introspection as a methodology in linguistics. Plenary lecture at the
10th International Cognitive Linguistics Conference, July 2007,
Kraków.
Tummers, Jose, Kris Heylen and Dirk Geeraerts
2005 Usage-based approaches in Cognitive Linguistics: A technical state
of the art. Corpus Linguistics and Linguistic Theory 1: 225–61.
Xiao, Richard and Tony McEnery
2006 Collocation, semantic prosody, and near-synonymy: A cross-
linguistic perspective. Applied Linguistics 27: 103–29.
Empirical cognitive semantics: Some thoughts

Anatol Stefanowitsch

Abstract
The transformation of a discipline from an art to a science involves at least three
steps: First, the discipline must adopt the protocols and practices of empirical re-
search; second, it must adapt those protocols to its object of research and, in the
process of doing so, operationalize its theoretical concepts (i.e., redefine them in
terms that allow them to be measured objectively and reliably); third, it must rel-
egate to the metaphysical level all concepts that resist such a redefinition until an
operationalization has been found. This paper sketches out the current progress of
cognitive semantics along this three-step process and discusses different definitions
of the notion meaning – meaning as concept, as proposition, as reference and as
context of use – and their potential for operationalization. It is argued that while
the field as a whole still seems hesitant about the future direction of the discipline,
there is a range of interesting methods that are available for transforming cognitive
semantics into empirical cognitive semantics.
Keywords: methodology, meaning, operationalization, philosophy of science

1. Introduction

A discipline attempting to transform itself from an art to a science must

take at least three steps, each of which will irrevocably and fundamentally
change the basic outlook of the research community and each of which will
therefore alienate a significant proportion of its members.1
The first step consists in adopting the protocols of empirical research
and the techniques to put them into practice. In the case of cognitive se-
mantics (or of Cognitive Linguistics more generally), the protocols and
techniques will be those already adopted by empirically-oriented research-
ers in psycholinguistics, corpus linguistics, and, arguably, discourse analy-
sis.
The second step consists in adapting these protocols and techniques to
the object of research specific to the discipline in question. The task for
researchers during this phase is to redefine all of the discipline’s theoretical
concepts in terms of a set of operations that must be performed to identify
356 Anatol Stefanowitsch

and measure the phenomena behind these concepts. This process of oper-
ationalization is a difficult undertaking even in the physical domain; it is all
the more difficult for disciplines dealing with intangibles such as linguistic
meaning.
The third, and perhaps most painful step does not involve adding some-
thing to the discipline in question, but in taking something away. It consists
in giving up those theoretical concepts that resist definition in the second
step, as well as those research questions that are dependent on these con-
cepts. In cognitive semantics, this may well mean giving up those central
theoretical concepts that define the field. The concepts in question may not
have to be given up permanently, of course, as an operational definition
could become available at a later point in time. The question must therefore
be answered, what place the concepts in question can be given within the
discipline until that point.
While individual researchers in the field of Cognitive Linguistics have
been working on the second of these three steps for some time now (the
contributions in the current volume document some of these attempts, as
do, for example, some of the contributions in Kristiansen et al. 2006, Gries
and Stefanowitsch 2006, Stefanowitsch and Gries 2006, Gonzales-Marques
et al. 2007, Zeschel 2008, and Brdar et al. forthc.). But the field as a whole
is currently, at best, taking the first of these steps – slowly, hesitantly, and
certainly not in unison.
This paper is intended as a contribution to this process. In the first sec-
tion, I will briefly sketch out the first of the three steps mentioned above,
i.e. the way in which, from my perspective, empirical methods entered the
field of Cognitive Linguistics and how they are currently evolving. In the
second section, I will then turn in more detail to the second step, that of
operationalization, and discuss the problems that this step involves for the
cognitively-oriented study of meaning. In the concluding section I will
briefly return to the consequences that the third step mentioned above, that
of discarding notions that cannot be operationalized, has for cognitive se-
mantics.

2. Methods in cognitive semantics

None of the authors who are, rightly or wrongly, regarded as having initi-
ated the field of Cognitive Linguistics are empirical researchers. Most of
them were originally trained in a framework whose primary method con-
 Empirical cognitive semantics 357

sists in (syntactic) argumentation on the basis of acceptability judgments on

constructed examples; a framework, moreover, that was built around the
assumption that linguistic analysis of any kind provided a direct window
into the human mind. It is not surprising, therefore, that all these authors
developed extensive theories about various aspects of human cognition
based on fairly traditional linguistic argumentation. Langacker is perhaps
most explicit about this. He refers to his Foundations of Cognitive Gram-
mar as “an exercise in speculative psychology” and he summarizes his
analytical procedure as follows:
I have adhered rather strictly, in developing my proposals, to the dictates of
both psychological plausibility and linguistic necessity; I have relied almost
exclusively on seemingly indisputable cognitive abilities (e.g. the ability to
compare two events and register a discrepancy between them), and I invoke
them for linguistic constructs that must somehow be be accommodated in
any comprehensive and explicit analysis (Langacker 1987: 6).
Other authors are less up-front about the speculative nature of their work,
but the complete absence of systematically collected empirical data in their
writings speaks for itself.2
The fact that the key claims of Cognitive Linguistics are claims about
human cognition rather than about language (or at least claims about lan-
guage in the context of cognition) made it inevitable, however, that some-
one would come along and subject these claims to empirical analysis. From
the early 1990s onward, a number of second-generation cognitive linguists
with a background in psychology began to introduce experimental methods
in Cognitive Linguistics research (Gibbs 1994 and the earlier studies he
summarizes there can be regarded as a milestone in this regard), and from
the mid-to-late 1990s, researchers began adding corpus-methods to the
Cognitive Linguistics toolbox (with Geeraerts et al. 1994 and Geeraerts
1997 as an important – perhaps the crucial – step on the way).
I will leave extensive discussion of the current state of empirical meth-
odology in Cognitive Linguistics to others for now (cf. Tummers et al.
2005 for an extensive discussion of the current state of cognitive corpus
linguistics, see also Stefanowitsch, in press) and selectively discuss their
potential to study “meaning” in the next two sections. First, let me briefly
dwell on some general problems of empirical methodology in Cognitive
Linguistics.
The introduction of experimental and corpus-based methods into a field
with very different methodological traditions is a dramatic – perhaps even
catastrophic – occurrence. It seems to me that two general reactions can be
358 Anatol Stefanowitsch

discerned within the Cognitive Linguistics community: denial and

overenthusiastic acceptance. The denial is reflected in the fact that many
prominent members of the Cognitive Linguistics community remain un-
convinced that empirical research is more than a convenient add-on, when
it confirms widely-held assumptions, or a nuisance, when it calls these as-
sumptions into doubt (at the expense of weakening my argument, I will
refrain from naming specific examples of this attitude here, as my intention
is not to criticize particular researchers but to observe the prevalence of a
general attitude). The overenthusiastic acceptance is reflected in the fact
that empirical methods are adopted by researchers with little or no back-
ground in the philosophy of science, the nuts and bolts of research design
or the statistical methods necessary to make sense of empirical results
(again, I will refrain from singling out individual researchers).
Both indifference and overenthusiastic acceptance make it difficult to
establish empirical thought as a valid addition to the Cognitive Linguistics
discourse. The fact that many researchers apply empirical methods incom-
pletely, incorrectly and without attention to established protocols of scien-
tific research produces results that are trivial, at best, and highly mislead-
ing, at worst. It thus reinforces the impression that empirical methods have
no genuine insights to add to theory building in Cognitive Linguistics.
Let me now turn to the question of operationalization to see whether this
impression is justified in principle and how it might be remedied.

3. Cognitive semantics and operational definitions

3.1. Operationalization

The adaptation of a set of research methods (protocols, procedures and

tools) to a new discipline may seem to be a minor technical issue, but it is,
in fact, the most drastic development that a field of research can undergo.
The reason is that these research methods cannot be used directly to meas-
ure the concepts used within the discipline to describe and explain its object
of study. Instead, the concepts themselves must be redefined in terms of
operations involving these research methods. This process of operationali-
zation fundamentally changes the nature of the concepts in question. Percy
Bridgman, one of the main proponents of operationalism, discusses these
consequences using the concept length:
 Empirical cognitive semantics 359

We evidently know what we mean by length if we can tell what the length
of any and every object is, and for the physicist nothing more is required.
To find the length of an object, we have to perform certain physical oper-
ations. The concept of length is therefore fixed when the operations by
which length is measured are fixed: that is, the concept of length involves as
much as and nothing more than the set of operations by which length is de-
termined. In general, we mean by any concept nothing more than a set of
operations; the concept is synonymous with a corresponding set of oper-
ations. (Bridgman 1991[1927]: 5, my emphasis).
Under this view, which is sometimes referred to as ‘strict’ operation-
alism, length is no longer regarded as a property of the world that is meas-
ured by a specific procedure involving devices like rulers, tape measures,
laser rangefinders, ultra-high frequency transmitters, etc. Instead, length is
simply the result of such a procedure. The procedures differ, of course: a
ruler must be placed in physical contact to the thing whose length it is sup-
posed to measure, UHF transmitters produce signals that must be triangu-
lated in order to produce a measure of distance, etc. Also, different ways of
using the same device may yield different results. Note that under a strict
operational view, there no longer is just one notion of length. Instead, there
are now at least four such notions: lengthRULER, lengthTAPEMEASURE, leng-
thRANGEFINDER, and lengthUHF, and their relationship to each other and to the
pre-operationalization notion length has become a highly complex issue in
its own right.
This sounds problematic enough, but there is an additional problem that
emerges as soon as we leave the physical domain, which we will clearly be
forced to do in the study of language, as there are relatively few properties
of language that can be physically measured. Acoustic properties of the
speech signal come to mind, and perhaps also articulatory mechanisms by
which the speech signal is produced. But as soon as we move beyond the
speech signal itself, things become less clear.
Bridgman’s example is quite suitable to demonstrating this (although
length may not among the most relevant phenomena in cognitive semantics,
it does play an important part, for example, in the study of information
structure). We could physically measure the length of (part of) a given ut-
terance in milliseconds, for example by recording it on magnetic tape and
then using the acoustic signal to trigger an appropriately wired stopwatch.
This would give us a definition of length as lengthMS.STOPWATCH, which
would be dependent on the accuracy of the triggering device and the stop-
watch itself. Or we could physically measure it by recording it digitally on
360 Anatol Stefanowitsch

a computer and then deriving its length from the segment of the data file
that represents the signal. This would give us the measure lengthMS.DIGITAL,
which would be dependent on the accuracy of the computer, the resolution
of the digital file, etc. The difference between these two methods corres-
ponds to the difference between lengthRULER and lengthRANGEFINDER. The
additional problem emerges if we want to measure the length of the same
utterance in terms of theoretically relevant units such as number of phon-
emes (lengthPHONEMIC), number of graphemes (lengthGRAPHEMIC), number of
morphemes (lengthMORPHEMIC), number of syllables (lengthSYLLABIC), number
of words (lengthLEXEMIC), etc. The first problem is that none of these can be
regarded self-evidently to the “correct” way of measuring the length of an
utterance. These measures may produce conflicting results such that A may
be longer than B according to one of these measures and B longer than A
according to another. The researcher must therefore decide which of these
new notions of “length” is plausibly relevant in the context of their investi-
gation. The second problem is that even once we decide on one of these
notions of length, the process of measuring it is highly complex. To meas-
ure lengthPHONEMIC, for example, we need something fundamentally differ-
ent from a tape recorder and a watch, namely, agreed-upon and replicable
procedures for determining what counts as a phoneme in a given language,
for identifying phoneme boundaries in actual words and utterances, and for
counting the segments in between the boundaries. Each of these procedures
is highly theory dependent.
This does not constitute an insurmountable obstacle from the perspec-
tive of operationalism – even ‘strict’ operationalism is not limited to phys-
ical measuring devices:
If the concept is physical, as of length, the operations are actual physical
operations, namely, those by which length is measured; or if the concept is
mental, as of mathematical continuity, the operations are mental operations,
namely those by which we determine whether a given aggregate of magni-
tudes is continuous. It is not intended to imply that there is a hard and fast
division between physical and mental concepts, or that one kind of concept
does not always contain an element of the other; this classification of con-
cept is not important for our future considerations. (Bridgman 1991[1927]: 5).
Although Bridgman is thinking of mathematics when he talks about “men-
tal operations”, it is clear that the nature of the “mental operations” in-
volved in the operationalization of some phenomenon may extend beyond
mathematics, as long as they are reliably replicable. In our example, struc-
tural phonemic analysis would be a mental operation that can be suffi-
 Empirical cognitive semantics 361

ciently explicitly described to be replicable, and hence, it can form part of

the definition of the lengthPHONEMIC of an utterance.

3.2. Four kinds of meaning

Although length is an important concept in linguistics and the problem of

coming up with a useful operationalization is a real one (and one that has
not been solved conclusively), it is almost trivial compared to the problems
that a truly empirical semantics would face in operationalizing its core
theoretical concepts.
The problem that will be discussed in the remainder of this paper is the
first one facing the discipline: that of defining the notion meaning itself.
Broadly speaking, there are four ways in which the term meaning has been
defined in the Cognitive Linguistics literature: (i) meaning as concept, (ii)
meaning as proposition, (iii) meaning as reference and (iv) meaning as
context of use. I will discuss each of these definitions in turn, not in terms
of which of them best reflects a cognitive approach to semantics, although
this is certainly a question that should be central to a cognitive semantics
research program, but in terms of how they can be defined operationally. At
first glance, all of them seem difficult, perhaps impossible, to operational-
ize, but in fact, there are methods in linguistics that can be, and have been,
used to measure, and hence to operationalize, the phenomena captured by
these definitions.

3.2.1. Meaning as concept

The dominant notion of meaning in Cognitive Linguistics is captured by the

following quotation from Langacker’s Foundations of Grammar, but simi-
lar notions underlie work by Lakoff (1987), Fauconnier (1995), and others:
I take it as self-evident that meaning is a cognitive phenomenon and must
eventually be analyzed as such. Cognitive grammar therefore equates mean-
ing with conceptualization (explicated as cognitive processing). In doing so,
it conflicts with major traditions of semantic theory ... in particular with the
many varieties of formal semantics based on truth-conditions, as well as the
newer “situation semantics” ... (Langacker 1987: 5)
Langacker takes a fairly abstract perspective on conceptualization/cognitive
processing; he takes great care to point out that a concept in his theory is
362 Anatol Stefanowitsch

not to be regarded as a “mental picture” (Langacker 1987: 97) and although

he makes heavy use of analogies to visual processing throughout his work,
it is clear that these really are meant to be just that – analogies –, and that it
is the abstract structure of these analogies that he is interested in.
It is a central tenet of Cognitive Linguistics that language should be
viewed, first and foremost, as a cognitive/conceptual phenomenon. With
his view of meaning, Langacker simply takes the implications of this tenet
to their logical conclusion. In doing so, however, he engages in a far-
reaching definitional act: he is not merely saying that conceptualization
exists, that it is relevant to language or that it is useful in explicating the
pre-scientific notion meaning; instead, he is saying that the word meaning
should be understood to refer exclusively to “conceptualization (explicated
as cognitive processing)”. He also explicitly rejects all other definitions of
the term: while he would not, presumably, deny that truth conditions
(whether applied to possible worlds or to situations) exist (i.e., that they can
be stated by a human observer), he states quite unambiguously that they are
irrelevant to an explication of the term meaning.
The idea that cognitive processing is relevant to the scientific investiga-
tion of language is no longer particularly controversial: most (perhaps all)
psycholinguistic approaches to language involve an attempt to measure
“cognitive processing” and there is, by now, a large number of standard
experimental protocols suitable for this task. These protocols themselves
can be regarded as operational definitions of whatever aspect of cognitive
processing they are meant to measure. They often involve measuring reac-
tion times, for example in the context of priming experiments or self-paced
reading/listening tasks. Thus, they define cognitive processing as “what
happens in a speaker’s head between the time they perceive a prime or a
stimulus and the time that they react”, e.g. by pushing a button. The content
of this processing activity is not measured but inferred, for example, by
comparing reaction times to different kinds of stimuli or to the same stimu-
lus in different contexts.
Clearly this looks like an extremely impoverished definition of process-
ing if you compare it to discussions in the Cognitive Linguistics literature,
for example, of “mental spaces” (Fauconnier 1985), “image schemas” and
“image schema transformations” (Lakoff 1987) or “construal” (Langacker
1987, 1991), to name just a few. However, by varying stimuli and response
types, psycholinguists have used this seemingly simple method to gain a
impressively deep knowledge about many aspects of cognitive processing,
 Empirical cognitive semantics 363

ranging from the time course of lexical access to the way in which syntactic
ambiguities are resolved during parsing.
A number of researchers have applied these standard protocols to re-
search issues of interest to cognitive linguists, such as the structure of lexi-
cal categories (e.g. Rosch and Mervis 1975), metaphor (Kemper 1989),
metonymy (Gerrig 1989) or indirect speech acts (Gibbs 1986), and their
work has been extended in many directions in Cognitive Linguistics.
There is one strand of research in the field of Cognitive Linguistics that
can be more closely related to the Langackerian definition of meaning: the
work on “simulation semantics”. Simulation semantics is characterized as
follows:
The study of how different aspects of language contribute to the construc-
tion of mental imagery, and the corresponding theory of linguistic meaning
as linguistic specifications of what and how to simulate in response to lan-
guage, is known as simulation semantics. (Bergen 2007: 278)
The definition of meaning as “specifications of what and how to simulate”
sounds like a reformulation of Langacker’s notion of conceptualization in
terms that are more amenable to operationalization – essentially, all that is
now needed is a set of procedures to measure mental simulation. The set of
procedures actually used by proponents of simulation semantics is con-
strained additionally by the assumption that simulation is embodied, i.e.
that it makes use of perceptual and motor circuitry in the brain:
As for the actual mechanisms underlying language processing on this em-
bodied view, understanding a piece of language is hypothesized to entail
performing mental perceptual and motor simulations of its content ... This
implies that the meanings of words and of their grammatical configurations
are precisely the contributions those linguistic elements make to the con-
struction of mental simulations (Bergen 2007: 278)
Measurements for mental simulation thus include neuro-imaging tech-
niques in addition to psycholinguistic techniques; all these techniques are
employed in a way that is suitable to measuring the activation of particular
motor or perceptual circuits in the brain in reaction to linguistic stimuli (see
the overview in Bergen 2007).
Simulation semantics is not the only way to operationalize a definition
of meaning as “conceptualization (explicated as cognitive processing)”, but
it shows in an exemplary manner the steps necessary to turn theoretical
notions into measurements. First, the pre-scientific notion meaning is re-
defined as cognitive processing; this brings it into the domain of the poten-
364 Anatol Stefanowitsch

tially measurable, but it does not yet constitute an operational definition.

Next, cognitive processing is explicated as mental simulation, with the
assumption that simulation in response to language uses the same neural
structures also used in perceiving and acting in the world. Other explica-
tions of cognitive processing might do without this assumption (see next
section), but making this assumption paves the way for the final step: men-
tal simulation is measured, for example, by observing blood flow in one
particular area of the brain (neuroimaging) or by priming experiments (psy-
cholinguistics). From the point of view or scientific modeling, these meas-
urements now constitute a definition of meaning, and each way of measur-
ing is its own definition (meaningNEURAL.ACTIVATION, meaningPRIMING, etc.).
I am not aware of any corpus-linguistic operationalizations of Lan-
gacker’s notion of meaning as cognitive processing, nor is it clear to me
what such an operationalization might look like. This is not to say that cor-
pus linguistics cannot contribute to the study of conceptualization in Lan-
gacker’s sense, but it can only do so much more indirectly than the
psycholinguistic methods discussed above (much like off-line psycholin-
guistic techniques like sorting tasks/similarity judgments, acceptability
judgments, etc.). The logic behind such studies is always the same: the
researcher observes some linguistic or language-related behavior and postu-
lates a set of cognitive processes that are consistent with this result. This
may happen in the context of inductive studies, where the result is observed
first and the cognitive processes are hypothesized afterwards, or in the con-
text of deductive studies, where the researcher predicts a particular result
based on a theoretical model of cognitive processing. Note that this logic is
not fundamentally different from the one behind the reaction-time studies
discussed above. Off-line studies, amongst which corpus studies may be
counted, generally differ from on-line studies in that the subjects’ behavior
(in this case, their language production) is more complex and temporally
further removed from the cognitive processes that presumably gave rise to
them. In the case of corpus studies, the (verbal) behavior of subjects is, in
addition, produced in natural, uncontrolled settings. Realistically, such
results can therefore only distinguish relatively broad classes of theoretical
models, nowhere near the resolution necessary to address many of the de-
tailed proposals found in the Cognitive Linguistics literature; crucially, in
the context of the present discussion, they can serve as operationalizations
of cognitive processing only to the extent that the operations necessary to
get from these results to a statement about the cognitive process that pro-
duced (or contributed to) these results is laid out in sufficient detail to be
 Empirical cognitive semantics 365

reliably replicable. On the whole, corpus linguists are remarkably lax in

laying out such operations explicitly, given their rigorous attention to other
aspects of scientific research (such as, for example, the statistical modeling
of their results). Studies that take important steps towards remedying this
are, for example, Wulff (2003, 2009).

3.2.2. Meaning as proposition

Langacker’s rejection of truth conditional semantics and situation seman-

tics, widely shared among cognitive linguists, presumably extends to all
propositional approaches to meaning. However, if we interpret the term
‘cognitive semantics’ a little more broadly, there are researchers within this
research program who explicitly adopt a propositional view of meaning.
Among those, Wierzbicka’s Natural Semantic Metalanguage (NSM) is
perhaps the most notable. NSM posits the existence of “universal concep-
tual primitives” that are combined according to the rules of a “universal
syntax of meaning” (Wierzbicka 1996: 20) and both are located quite ex-
plicitly in the human mind:
the theory posits the existence not only of an innate and universal “lexicon
of human thoughts”, but also of an innate and universal “syntax of human
thoughts”. Taken together, these two hypotheses amount to positing some-
thing that can be called “a language of thought”, or lingua mentalis ...
(Wierzbicka 1996: 20).
There are other authors in Cognitive Linguistics that make use of proposi-
tional notation systems (for example, Talmy 2000), but as far as I can tell
they do not commit themselves to the claim that meaning actually is propo-
sitional in nature.
NSM is a radical redefinition of the pre-scientific notion meaning, but it
is not, of course, an operational definition – universal conceptual primitives
and a universal syntax of meaning are purely theoretical notions and no
explicit and replicable method exists by which they could be measured.
Wierzbicka sees herself as an empirical researcher, but her description of
her method is limited to statements like “conceptual primitives can be
found through in-depth analysis of any natural language” (Wierzbicka
1996: 13). Her actual method consists in explicating her intuitions about the
meaning of words and constructions in terms of the conceptual primitives
she has posited a priori – there is nothing in this procedure that could serve
as a basis for an operational definition.
366 Anatol Stefanowitsch

There have been attempts to explicate the (primitive) elements from

which NSM’s propositional structures are built up: Goddard (2001) has
attempted to redefine the notion semantic primitive in terms of “lexical
universals” – meanings that are lexified in all languages. Clearly such an
approach requires additional steps to count as an actual operationalization
(for example, a language-independent procedure for determining whether a
particular word has a particular meaning in a given language). Goddard’s
own work does not offer anything approaching such procedures, but by
redefining semantic primitives as lexical universals he has taken a crucial
step in bringing them into the domain of the empirically testable. As in the
case of the research on mental simulation discussed in the preceding sec-
tion, however, he has also fundamentally changed the ontological and epis-
temological status of the notion he set out to define.
It may be worthwhile, in the context of feature-based and/or proposi-
tional definitions of meaning to recall early psycholinguistic attempts to
identify semantic features, for example, by sorting tasks (e.g. Miller 1969)
as well as early work on the propositional nature of text representation (cf.
Kintsch 1974). The problem with such approaches is that they must be able
to distinguish feature-based/propositional representations from representa-
tions based on mental simulation (see preceding section). For example,
Miller (1969) shows that words that arguably share semantic features are
grouped together in sorting tasks. However, the same effect would emerge
if subjects mentally simulated the actions referred to by these verbs and
then grouped them on shared aspects of these simulations. Similarly,
Kintsch and Keenan (1973) show that reading rates and retention of senten-
ces depend on the number of propositions attributed to these sentences.
However, two sentences to which we can plausibly attribute different num-
bers of propositions presumably also differ with respect to the complexity
of the simulations they trigger, which means the latter could provide an
alternative explanation. Still, it does not seem inconceivable to devise ex-
periments to distinguish between propositional and simulated representa-
tions, and therefore both meaningSIMULATION and meaningPROPOSITION can be
regarded as potentially plausible operational definitions within cognitive
semantics.
 Empirical cognitive semantics 367

3.2.3. Meaning as reference

Given the strong focus on concepts and conceptualization found in Cogni-

tive Linguistics, it is perhaps not surprising that the real-world aspect of
meaning (Frege’s reference) has not received much attention in the field,
either in its own right or as a means of measuring conceptualization (al-
though a few examples, such as Lehrer 1982 and Schmid 1983 can be
found; also, arguably, one could classify some of the literature on semantic
networks, such as Lindner 1982 and Geeraerts 1983 as reference-oriented).
The general rarity of the meaning-as-reference approach is somewhat
surprising, since some of the foundational studies in cognitive semantics
rely heavily on naming and classification tasks which directly measure the
referential range of words (cf. the early work on prototypes [Rosch 1973,
Rosch and Mervis 1975, Rosch et al. 1976], on variation in referential
boundaries [Labov 1973], on color categorization [Berlin and Kay 1969], to
name just a few). These studies are usually careful to point out that they are
not measuring real-world reference, but reference within a model of the
world that filtered through species-specific perceptual processes (cf. the
discussion in Rosch et al. 1978: 28ff.). Still, the fact remains that they de-
fine the meaning of words via properties of their referents, for example,
when Rosch et al.
demonstrat[e] that in taxonomies of common concrete nouns in English
based on class inclusion, basic objects are the most inclusive categories
whose members: (a) possess significant numbers of attributes in common,
(b) have motor programs which are similar to one another, (c) have similar
shapes, and (d) can be identified from averaged shapes of members of the
class (Rosch et al. 1976: 382).
There is even a corpus-linguistic operationalization of the referential theory
of meaning, pioneered by (and, as far as I can tell, limited to) Geeraerts et
al. (1994, cf. also Geeraerts 1997). Geeraerts and his colleagues study the
semantic development of the loanword legging(s) in Dutch by searching for
images of leg wear whose description includes the word legging(s) in
women’s magazines and mail-order catalogues published between 1988
and 1991. By classifying the depicted leg wear in terms of length, tightness,
the presence of a crease, material, function and sex of intended wearer and
then counting the frequency of these features and tracking this frequency
across time, they are able to identify a prototypical referent for the word as
well as changes in what speakers of Dutch considered to be such a proto-
type.
368 Anatol Stefanowitsch

Like the naming and classification tasks used by psycholinguists, Geer-

aerts’ method is so simple and elegant that one might have expected it to be
adopted widely as one of the central methods in prototype semantics, not
just in the context of diachronic research questions but also for determining
the synchronic structure of lexical categories. The constraint pointed out
above for psycholinguistic uses of referents also holds for the corpus-
linguistic version, but within this constraint, referential range is such a
straightforward operationalization of meaning that it hardly needs justifica-
tion. That it has not been used more widely in prototype research and other
areas of cognitive semantics is unlikely to be due to theoretically motivated
reservations; it is more likely that practical considerations make it unattrac-
tive: collecting suitable samples of images with descriptions is very time-
consuming. If this is the reason, then perhaps the wide availability of se-
mantically tagged images through image search engines such as Google
Images or Yahoo! Images and photo sharing sites like Picasa and Flickr
will inspire a new wave of reference-based research in cognitive semantics.
Referential definitions of meaning, whether one bases them on experi-
mental elicitation or on naturally occurring usage data, are obviously lim-
ited to the study of words (and other units of linguistic structure) that refer
to perceptible entities and situations, but this is largely a practical limita-
tion. For such units, meaningREFERENCE is such an obvious way of operation-
alizing meaning that the relatively minor role it plays in current cognitive
semantics is rather surprising.

3.2.4. Meaning as context of use

There are traditions in linguistics that look for the meaning of a linguistic
expression in the contexts in which they are used. Generally speaking,
Wittgenstein’s idea that “the meaning of a word is its use in the language”
(Wittgenstein, Philosophical Investigations, Sect. 43) can be seen as the
most general statement of this position, another one would be the assump-
tion in Conversation Analysis that meaning is negotiated by speakers in
linguistic interaction and that one can analyze meaning by analyzing clues
in the interactional context (for example, by observing how speakers hold
each other “accountable” for particular interpretations of what is said).
What I have in mind more specifically, though, is the idea that the
meaning of a word is reflected in the linguistic contexts in which it occurs
frequently. As far as I know, none of the leading theoreticians in Cognitive
 Empirical cognitive semantics 369

Linguistics have paid much attention to this idea; but the assumption that
the meaning of a word can be read off from other words that it frequently
occurs with is quite common in corpus linguistics. It goes back at least to
Firth, who argued that we can “know a word by the company it keeps”:
[A] text in such established usage may contain sentences such as ‘Don’t be
such an ass!’, ‘You silly ass!’, ‘What an ass he is!’ In these examples, the
word ass is in familiar and habitual company, commonly collocated with
you silly-, he is a silly-, don’t be such an-. You shall know a word by the
company it keeps! One of the meanings of ass is its habitual collocation
with such other words as those above quoted (Firth 1957: 11).
Firth’s example shows that he takes “linguistic context” to include not just
collocates (i.e., words) but also complex expressions (i.e., what has been
called patterns in Pattern Grammar, cf. Hunston and Francis 1999; or col-
lostructions in Collostructional Analysis, cf. Stefanowitsch and Gries
2003). Firth’s example is concerned with distinguishing between different
senses of a word, and this is one of the most frequent applications to which
Firth’s method has been put (another is differentiating the meaning of near-
synonyms, cf. e.g. Kennedy 1991, Church et al. 1991, Stefanowitsch and
Gries 2003, Gries and Stefanowitsch 2004).
A number of researchers, beginning with Atkins (1987), have recently
taken up Firth’s idea of identifying for a given word some well-defined set
of lexical or lexico-grammatical co-occurrence patterns and using that set
to determine the meaning of that word in general (cf. e.g. Stubbs’ [1995]
notion of a “semantic profile” or Lowe’s [1997] “distributional profiles”) or
to distinguish different senses of the word in question (cf. Hanks’ [1996,
1997] notion of a “behavioral profile”).
Researchers differ in how much and what type of information they in-
clude in their co-occurrence profiles. Typically, they restrict themselves to
formal (morphological and syntactic) properties of the context, but Gries’
(2006) explicitly cognitive linguistic and rigorously quantified variant (for
which he uses Hanks’ term behavioral profile) also includes semantic
properties of the words themselves (e.g. different verb senses, the process
type of verbs, the animacy of subjects and objects, etc., cf. also Gries and
Divjak 2009; this vol.).
As far as I can tell, all of the researchers cited here assume that co-
occurrence profiles reflect the meaning (or different senses) of a word, but
none of them suggest that these profiles are the meaning of that word (as
Firth did). This seems to be a small step conceptually, though, and it would
turn the notion of a co-occurrence profile into an operational definition of
370 Anatol Stefanowitsch

meaning. Of course, in order to use co-occurrence profiles of any kind as

operational definitions for meaning, they must exclude semantic informa-
tion and be strictly limited to formal contextual features (as in Janda and
Solovyev’s [2009] “constructional profiles”).
“Meaning as context of use” is an interesting operational definition of
meaning in that it provides a link between a well-established corpus-
linguistic research tradition and Langacker’s idea that linguistic representa-
tions emerge from linguistic usage. It is also the only possible genuinely
corpus-based definition of meaning. Of course, as with the other oper-
ational definitions discussed here, it changes the meaning of the every-day
word meaning quite radically. This raises the question how all these differ-
ent definitions are related to each other, a question to which I will turn in
the next subsection.

3.3. Theory and operationalization

If the field of cognitive semantics were to adopt operationalism along the

lines sketched out above, the notion meaning would be replaced by notions
like meaningMENTAL.SIMULATION, meaningCOOCCURRENCE.PROFILE, meaningLEXI-
CAL.UNIVERSALS, and meaningREFERENCE. Actually, even this is a simplification,
as each of these notions would in turn be replaced by specific operational
procedures: mental simulation can be measured in different ways and each
of them would constitute an operational definition in its own right, giving
us, for example, meaningMENTAL.SIMULATION-FMRI, and meaningMEN-
TAL.SIMULATION-PRIMING; similarly, each specific suggestion as to how co-
occurrence profiles should be gathered and what information should be
included in them would give us a different operational definition (e.g.
meaningPROFILE-CONSTRUCTIONAL, meaningPROFILE-BEHAVIORAL-HANKS, meaning-
PROFILE-BEHAVIORAL-GRIES, etc.); a referential view could give us meaningREFER-
ENCE-PICTURE.NAMING, meaningREFERENCE-CATALOGUE, etc., and so on.
Faced with this multitude of specific definitions which replace what was
traditionally thought of as a unified, if extremely ill-defined, theoretical
notion, we could ask ourselves how we should think of their relationship to
each other. There are two possible answers to this question, a moderate one
and a radical one.
The moderate answer is that operational definitions cannot be all there is
to a discipline, and that the operational definitions must be held together by
the overarching notion that inspired them in the first place. In the case of
 Empirical cognitive semantics 371

empirical cognitive semantics, researchers would continue to assume that

there is such a phenomenon as meaning that is separate from the various
operational definitions and that every operational definition is simply an
attempt to capture (some aspect of) this phenomenon.
This answer is initially attractive for several reasons. First, strict oper-
ationalism has often been criticized as being irreconcilable with scientific
progress, as every change in an operational procedure would have to be
treated as a change of the object of study (Boyd 1991: 9). Thus, no proced-
ure could ever be improved. Second, it has been pointed out that in actual
practice, researchers do “treat the new procedures as improvements or ex-
tensions of the old ones”, and thus behave
exactly as they should if the new laboratory procedures were improved or
extended procedures for measuring, detecting, or otherwise assessing a pre-
viously studied unobservable feature of the world. (Boyd 1991: 9, cf. also
Boyd 2008).
In other words, scientists do not seem to adhere to operationalism, but in-
stead seem to treat operational definitions as proxies for something else,
something that cannot be measured at all. Third, we seem to have direct
access to that something in the case of semantics: our own conscious ex-
perience of what it feels like to understand language. This experience itself
cannot be measured, but we may recognize intuitively that different oper-
ational definitions correspond to it more or less closely (I certainly do).
However, the fact that an answer is subjectively attractive or that it avoids
certain difficult issues in bringing together scientific practice and the phi-
losophy of science does not necessarily make it a good answer.
The second, radical answer to the question of how the different oper-
ational definitions are related to each other is simply, that this remains to be
demonstrated empirically. This answer accepts that there is no a priori rea-
son to believe that the various operational definitions are related to some
common, unmeasurable phenomenon. The task of the semanticist is to look
for correlations between the phenomena measured by the different oper-
ational procedures and to posit causal relationships where such correlations
are found. These causal relationships would be the core of a model of se-
mantics. This model would be unlikely to correspond particularly closely to
our conscious experience of understanding language, but whether or not a
scientific theory matches our conscious experience of the object of study is
not a measure of quality. As for the problem of changing operational pro-
cedures, I think it is fair to say that critics of strict operationalism overstate
372 Anatol Stefanowitsch

the issue. There is a difference between improving the accuracy of a meas-

urement, switching to a different measuring procedure that plausibly meas-
ures the same thing, and switching to a different procedure that measures
something else, and this difference can be demonstrated conceptually and
empirically by comparing a new procedure directly to an old one (which is
routinely done when operational procedures are changed).
The two answers are not mutually exclusive: it is possible to treat the re-
lationship between different operational definitions for what used to be
called “meaning” as an empirical issue (i.e., to look for correlations and
causal relations) and, at the same time, continue a sort of metaphysical
discourse about some unmeasurable phenomenon in the real world (the
structure of language, the mind of a speaker, etc.) that potentially unites all
of these measurements. In a sense, this combined practice is close to what
empirical cognitive semanticists are already doing, so it would seem to be a
reasonable first step.

4. Concluding remarks

While empirical work in cognitive semantics addresses a wide range of

issues that are relevant to a “cognitive semantics”, it seems clear that no
attempt has yet been made to formulate a comprehensive model of cogni-
tive semantics in such a way that it can be operationalized along the lines of
one or more of the approaches sketched out above.
No doubt this is partly due to the fact that empirically oriented research-
ers in cognitive semantics are still at the beginning of the second of the
three steps mentioned in the introduction. However, it is also due to the fact
that the specific theories of cognitive semantics that currently exist in the
field are both too general and too specific to be amenable to operationaliza-
tion. They are too general in that they make claims far beyond anything that
could ever be tested empirically, and that thus largely belong to what Pop-
per refers to as metaphysics (note that this use of the term diverges from the
philosophical mainstream). They are too specific in that they apply these
ideas to phenomena at a level of detail where they would be difficult to test
even if they were in principle testable.
Current cognitive linguistic theories of meaning have played an import-
ant role in the formation of the discipline. They have sketched out new
ways to look at meaning (or rediscovered forgotten ones) and have thus
inspired, and continue to inspire, empirical researchers such as the contri-
 Empirical cognitive semantics 373

butors to the present volume and others cited above. They should be recog-
nized as the corner stones of the emerging discipline of empirical cognitive
semantics, but we should be careful to keep them out of scientific models
of meaning in human languages and human minds.
This does not mean that we have to abandon them altogether – we could
argue, with Popper (1959: 16), that it might be useful to keep them around,
not because discussing them can be expected to advance directly our under-
standing of the object of study, but because even unfounded speculation
will occasionally advance novel ideas that may inspire additional oper-
ationalizations as the cycle of hypothesis testing (cf. Geeraerts, this vol.)
continues to move the scientific debate in the field forward.
However, keeping these theories around should not be mistaken as an
escape route for researchers unwilling to dirty their hands by delving into
empirical data; if every currently non-empirical author in the field were to
move their interests exclusively to contributing to metaphysical debates of
semantics, the discipline of Cognitive Linguistics would become so top-
heavy that it would be certain to topple, especially as it is unlikely that all
of them, or even the majority, would have much to contribute to meta-
physical debates.
Instead, the relegation of much of the current theoretical discourse in
cognitive semantics to the metaphysical level should inspire a growing
number of researchers to confine themselves to the research questions ad-
dressable by the currently available methods and techniques and to take up
the task of extending these methods and techniques as far as possible. This
would have two consequences.
First, Cognitive Linguistics will move away from its original axioms
and become, in many areas of research, indistinguishable from other ap-
proaches to the same areas. This has already happened in areas like applied
linguistics, corpus-linguistics and sociolinguistics. The main feature that
still makes work in these areas “cognitive” is the self-classification of the
researchers in question, and, perhaps, a more explicit discussion of how the
results of their research relates to human conceptualization and cognitive
processing (although non-cognitive researchers in these fields do, in many
cases, recognize the cognitive implications of their work quite explicitly,
see, for example, Hoey 2005). Second, cognitive semantics might ulti-
mately have to shift its focus away from some of the research questions that
are currently central to it, as it becomes clear that they are forever beyond
operationalization and measurement.
374 Anatol Stefanowitsch

Cognitive semantics, in other words, will first become empirical cogni-

tive semantics (the present volume contributes to this process), and then
simply empirical semantics. From the inside perspective of the current
Cognitive Linguistics community this may be undesirable, as it seems to
rob the discipline of its unique selling point. But from the perspective of a
future science of empirical semantics, the shedding of the adjective cogni-
tive is a minor loss, and it may even be a decisive step in the emancipation
of the discipline from its tradition as an exercise in speculative psychology.
To the extent that semantics is a cognitive phenomenon, the results of em-
pirical semantics will show this with or without the label cognitive.

Notes

1. This paper is loosely based on an impromptu talk I delivered at the workshop

on Empirical Cognitive Semantics at the 10th International Cognitive Lin-
guistics Conference in Kraków. I would like to thank the participants of this
workshop, especially Daniel Wiechmann and Dirk Geeraerts, for their stimu-
lating discussion, and Juliana Goschler for her comments on an early draft of
this paper. Finally, I would like to thank the organizer and editor of the pres-
ent volume, Kerstin Fischer, for her comments and her angelic patience in
waiting for me to finish the paper. None of them are responsible for my unfin-
ished ideas, premature claims and eclectic and incomplete references to the
literature.
2. The indifference towards empirical methods is also reflected in current Cog-
nitive Linguistics textbooks: Ungerer and Schmid (1996, 2006) and Croft and
Cruse (2004) do not contain any references to empirical methods at all; Lee
(2002) limits its empirical remarks, rather idiosyncratically, to a short chapter
on discourse analysis; Evans and Green (2006) introduce introspection as the
sole methodology of the discipline and claim that it is a “‘window’ to the
underlying system” (Evans and Green 2006: 17). This indifference is all the
more surprising as some of these authors are clearly aware of the existence of,
and need for, empirical methods: Schmid and Lee make extensive use of cor-
pus linguistic data in their own research and Evans has collaborated with a
number of psycholinguists.
 Empirical cognitive semantics 375

References

Atkins, Beryl T.
1987 Semantic-ID tags: Corpus evidence for dictionary senses. In The
Uses of Large Text Databases. Proceedings of the Third Annual
Conference of the New OED Centre. University of Waterloo,
Canada, 17–36.
Bergen, Benjamin
2007 Experimental methods for simulation semantics. In Methods in Cog-
nitive Linguistics, Monica Gonzalez-Marquez, Irene Mittleberg, Se-
anna Coulson and Michael J. Spivey (eds.), 277–301. Amsterdam /
Philadelphia: John Benjamins.
Berlin, Brent and Paul Kay
1969 Basic Color Terms. Their universality and evolution. Berkeley:
University of California Press.
Boyd, Richard
1991 Confirmation, semantics, and the interpretation of scientific theories.
In The Philosophy of Science, Richard Boyd, Philip Gasper and J.D.
Trout (eds.), 5–35. Cambridge, MA: MIT Press.
2008 Scientific realism. In The Stanford Encyclopedia of Philosophy (Fall
2008 Edition), Edward N. Zaltaed. (ed.). Online at
plato.stanford.edu/archives/fall2008/entries/scientific-realism/ (ac-
cessed 9th February, 2009).
Brdar, Mario, Milena Žic Fuchs and Stefan Th. Gries (eds.)
forthc. Expanding Cognitive Linguistic Horizons. Amsterdam / Philadel-
phia: John Benjamins.
Bridgman, Percy W.
1991 The operational character of scientific concepts. In The Philosophy
of Science, Richard Boyd, Philip Gasper and J.D. Trout (eds.), 57–
69. Cambridge: MA: MIT Press. [Originally Chapter 1 of Percy
Bridgman (1927), The Logic of Modern Physics, New York: Mac-
Millan.]
Church, Kenneth W., Patrick Hanks, Donald Hindle and William Gale
1991 Using statistics in lexical analysis. In Lexical Acquisition: Using on-
line resources to build a lexicon, Uri Zernik (ed.), 115–164. Mawah,
NJ: Erlbaum.
Croft, William and Alan Cruse
2004 Cognitive Linguistics. Cambridge: Cambridge University Press.
Evans, Vyvyan and Melanie Green
2006 Cognitive Linguistics. An introduction. Edinburgh: Edinburgh Uni-
versity Press.
376 Anatol Stefanowitsch

Fauconnier, Gilles
1985 Mental Spaces. Aspects of meaning construction in natural lan-
guage. Cambridge, MA: MIT Press.
Firth, John R.
1957 A synopsis of linguistic theory 1930–1955. In Studies in Linguistic
Analysis. (Special ed. Transactions of the Philological Society), J. R.
Firth (ed.), 1–32. Oxford: Basil Blackwell.
Geeraerts, Dirk
1983 Prototype theory and diachronic semantics. A case study. Indoger-
manische Forschungen 88: 1–32.
1997 Diachronic Prototype Semantics. A contribution to historical lexi-
cology. Oxford: Clarendon Press.
this vol. The doctor and the semantician.
Geeraerts, Dirk, Stefan Grondelaers and Peter Bakema
1994 The Structure of Lexical Variation. Meaning, naming, and context.
Berlin / New York: Mouton de Gruyter.
Gerrig, Richard J.
1989 The time-course of sense creation. Memory and Cognition 17: 194–
207.
Gibbs, Raymond W., Jr.
1986 What makes some indirect speech acts conventional. Journal of
Memory and Language 25: 181–196.
1994 The Poetics of Mind. Figurative thought, language, and understand-
ing. Cambridge: Cambridge University Press.
Goddard, Cliff
2001 Lexico-semantic universals: A critical overview. Linguistic Typol-
ogy 5(1): 1–66.
Gonzalez-Marquez, Monica, Irene Mittelberg, Seana Coulson and Michael J.
Spivey (eds.)
2007 Methods in Cognitive Linguistics. Amsterdam / Philadelphia: John
Benjamins.
Gries, Stefan Th.
2006 Corpus-based methods and Cognitive Semantics: The many mean-
ings of to run. In Corpora in Cognitive Linguistics. Corpus-based
approaches to syntax and lexis, Stefan Th. Gries and Anatol Ste-
fanowitsch (eds.), 57–99. Berlin / New York: Mouton de Gruyter.
Gries, Stefan Th. and Dagmar Divjak
2009 Behavioral profiles: A corpus-based approach to Cognitive Semantic
analysis. In New Directions in Cognitive Linguistics, Vyvyan Evans
and Stephanie S. Pourcel (eds.), 57–75. Amsterdam / Philadelphia:
John Benjamins.
 Empirical cognitive semantics 377

this vol. Quantitative approaches in usage-based Cognitive Semantics:

Myths, erroneous assumptions, and a proposal.
Gries, Stefan Th. and Anatol Stefanowitsch
2004 Extending collostructional analysis: A corpus-based perspective on
‘alternations’. International Journal of Corpus Linguistics 9(1): 97–
129.
Hanks, Patrick
1996 Contextual dependency and lexical sets. International Journal of
Corpus Linguistics 1(1): 75–98.
1997 Lexical sets: relevance and probability. In Translation and Meaning,
Part 4, Barbara Lewandowska-Tomaszczyk and Marcel Thelen
(eds.), 19–139. Maastricht: Hogeschool Maastricht, School of Trans-
lation and Interpreting.
Hoey, Michael
2005 Lexical Priming. A new theory of words and language. London:
Routledge.
Hunston, Susan and Gil Francis
1999 Pattern Grammar. A corpus-driven approach to the lexical grammar
of English. Amsterdam / Philadelphia: John Benjamins.
Janda, Laura A. and Valery D. Solovyev
2009 What constructional profiles reveal about synonymy: A case study
of Russian words for SADNESS and HAPPINESS. Cognitive Linguistics
20(2): 367–393.
Kemper, Susan
1989 Priming the comprehension of metaphors. Metaphor and Symbolic
Activity 4: 1–17.
Kennedy, Graeme
1991 Between and through: The company they keep and the functions
they serve. In English Corpus Linguistics, Karin Aijmer and Bengt
Altenberg (eds.), 95–110. London: Longman.
Kintsch, Walter
1974 The Representation of Meaning in Memory. Hillsdale: Erlbaum.
Kintsch, Walter and Janice Keenan
1973 Reading rate and retention as a function of the number of proposi-
tions in the base structure of sentences. Cognitive Psychology 5:
257–274.
Kristiansen, Gitte, Michel Achard, René Dirven and Francisco José Ruiz de Men-
doza Ibáñez (eds.)
2006 Cognitive Linguistics. Current applications and future perspectives.
Berlin / New York: Mouton de Gruyter.
378 Anatol Stefanowitsch

Labov, William
1973 The boundaries of words and their meanings. In New Ways of Ana-
lyzing Variation in English, Charles Bailey and Roger Shuy (eds.),
340–373. Washington: Georgetown University Press.
Lakoff, George
1987 Women, Fire, and Dangerous Things. What categories reveal about
the mind. Chicago: The University of Chicago Press.
Langacker, Ronald W.
1987 Foundations of Cognitive Grammar. Vol. 1. Theoretical prerequi-
sites. Stanford, CA: Stanford University Press.
1991 Foundations of Cognitive Grammar. Vol. 2. Descriptive application.
Stanford, CA: Stanford University Press.
Lee, David
2002 Cognitive Linguistics. An introduction. Oxford: Oxford University
Press.
Lehrer, Adriene
1982 Wine and Conversation. Bloomington: Indiana University Press.
Lindner, Susan J.
1982 What goes up doesn’t necessarily come down: The ins and outs of
opposites. In Papers from the Eighteenth Regional Meeting of the
Chicago Linguistic Society, 305-323. Chicago: Chicago Linguistic
Society.
Lowe, Will
1997 Meaning and the mental lexicon. In Fifteenth International Joint
Conference on Artificial Intelligence, Proceedings. Vol. 1. Martha
E. Pollack (ed.), 1092–1098. San Francisco: Morgan Kaufman.
Miller, George A.
1969 A psychological method to investigate verbal concepts. Journal of
Mathematical Psychology 6: 169–191.
Popper, Karl
1959 The Logic of Scientific Discovery. London: Hutchinson & Co.
Rosch, Eleanor
1973 Natural categories. Cognitive Psychology 4: 328–350.
1978 Principles of categorization. In Cognition and Categorization, El-
eanor Rosch and Barbara B. Lloyd (eds.), 27–48. Hillsdale:
Erlbaum.
Rosch, Eleanor and Carolyn B. Mervis
1975 Family resemblances: Studies in the internal structure of categories.
Cognitive Psychology 7: 573–605.
 Empirical cognitive semantics 379

Rosch, Eleanor, Carolyn B. Mervis, Wayne D. Gray, David M. Johnson and Penny
B. Braem
1976 Basic objects in natural categories. Cognitive Psychology 8(3): 382-
439.
Sandra, Dominiek and Sally Rice
1995 Network analyses of prepositional meanings: Mirroring whose mind
– the linguist’s or language user’s? Cognitive Linguistics 6(1): 89–
130.
Schmid, Hans-Jörg
1993 Cottage and Co., Idea, Start vs. Begin. Die Kategorisierung als
Grundprinzip einer differenzierten Bedeutungsbeschreibung. Tübin-
gen: Niemeyer.
Stefanowitsch, Anatol
in press Cognitive linguistics as a cognitive science. In Bi-directionality in
the Cognitive Sciences. Examining the interdisciplinary potential of
cognitive approaches in linguistics and literary studies, Marcus Cal-
lies, Astrid Lohöfer and Wolfram Keller (eds.). Amsterdam /
Philadelphia: John Benjamins.
Stefanowitsch, Anatol and Stefan Th. Gries
2003 Collostructions: Investigating the interaction of words and construc-
tions. International Journal of Corpus Linguistics 8(2): 209-243.
Stubbs, Michael
1995 Collocations and semantic profiles: On the cause of the trouble with
quantitative studies. Functions of Language 2(1): 23-55.
Talmy, Leonard
2000 Toward a Cognitive Semantics. Vol. 1. Concept structuring systems.
Vol. 2. Typology and process in concept structuring. Cambridge,
MA: MIT Press.
Ungerer Friedrich and Hans-Jörg Schmid
1996 An Introduction to Cognitive Linguistics. London: Longman.
2006 An Introduction to Cognitive Linguistics. Second edition. London:
Pearson Longman.
Wittgenstein, Ludwig
1953 Philosophische Untersuchungen. Philosophical investigations.
Trans. Gertrude E.M. Anscombe. Oxford: Blackwell.
Wierzbicka, Anna
1996 Semantics: Primes and universals. Oxford: Oxford University Press.
Wulff, Stefanie
2003 A multifactorial corpus analysis of adjective order in English. Inter-
national Journal of Corpus Linguistics 8(2): 245–82.
2009 Rethinking Idiomaticity. A usage-based approach. London / New
York: Continuum.
380 Anatol Stefanowitsch

Zeschel, Arne (ed.)

2008 Usage-based approaches to language processing and representa-
tion. (Special ed. Cognitive Linguistics 19). Berlin / New York:
Mouton de Gruyter.
Author index

Abbot-Smith, Kirsten, 204, 218 Boyes-Braem, Penny, 132, 200

Abraham, Werner, 305, 308, 327 Braem, Penny, 379
Achard, Michel, 31, 37, 43, 58, 77, Brdar, Mario, 352, 356, 375
96, 237, 301, 377 Brems, Lieselotte, 28
Andonova, Elena, 186, 196 Brennan, Robert L., 207, 219
Arppe, Antti, 15, 20, 23, 27, 342, Bresnan, Joan, 24, 28, 264–265, 296,
344–345, 350 299, 320, 327
Atkins, Beryl T., 19, 31, 45, 56, 180, Bridgman, Percy W., 358–360, 375
198, 338, 350, 369, 375 Broekhuis, Hans, 298–299
Brugman, Claudia M., 102, 128
Baayen, R. Harald, 27–28, 218, 265, Bryant, John, 181, 195–197, 199
299, 320, 325, 327 Buhusi, Catalin V., 306, 327
Baichi, Anna, 196, 200 Burgess, Curt, 225, 238
Bakema, Peter, 32, 45, 57, 116–117, Bybee, Joan, 23, 26, 28, 98, 126–
129, 265, 376 128, 201–204, 218–219, 312, 327
Baldwin, Timothy, 229, 236 Byloo, Pieter, 329
Bannard, Colin, 229, 236
Barkema, Henk, 233, 236 Cacciari, Cristina, 233, 236–238
Barlow, Michael, 58, 143, 152, 219, Campbell, Donald T., 67, 78
302 Caramazza, Alfonso, 116, 128, 130–
Bateman, John A., 188, 198 131
Bednarek, Monika, 44, 56 Carroll, John, 229, 238
Beliën, Maaike, 287, 299 Casella, George, 320, 329
Berck, Peter, 283, 300 Casenhiser, Devin, 280, 300
Berez, Andrea L., 335, 340, 350 Chang, Nancy, 141, 152, 179–182,
Bergen, Benjamin, 181–182, 196, 195–197, 199
363, 375 Chomsky, Noam, 13, 66, 76, 96
Berlin, Brent, 27–41, 45, 56–59, 76– Church, Kenneth W., 107, 128, 369,
78, 96–98, 129, 131–132, 176– 375
177, 199, 265–268, 301–303, Clancy, Steven J., 24, 28
329, 351, 353, 367, 375–377, 380 Claridge, Claudia, 95–96
Bermel, Neil, 306, 316, 327 Clark, Herbert H., 187, 193, 197
Berry-Rogghe, Godelieve L., 229– Clear, Jeremy, 107, 129
230, 233, 236 Colleman, Timothy 24, 28–29, 38,
Bi, Yanchao, 128, 379 54, 176, 271, 274–275, 281, 286,
Biber, Douglas, 180–181, 196–197 296, 299–300
Boeckh, August, 65, 76 Comrie, Bernard, 139, 152, 306,
Bolton, John, 237 309, 327
Boogaart, Ronny, 305, 307, 324, 327 Connine, Cynthia M, 225, 238
Boyd, Richard, 371, 375 Conrad, Randi, 197
382 Author index

Costa, Albert, 128, 130 Durkin, Kevin, 14, 30

Coulson, Seanna, 33, 35, 40, 43, 57, Dziwirek, Katarzyna, 21, 23, 26, 29,
266, 268, 375–376 30, 37, 351
Croft, William, 21, 28, 86, 96, 127,
129, 141–143, 146, 152, 227, Eddington, David, 23, 28, 201–202,
237, 374–375 218–219
Cruse, Alan, 143, 152, 227, 237, Elman, Jeff, 50, 56
374–375 Ericsson, K. Anders, 67, 76
Cueni, Anna, 28, 265, 299, 327 Erman, Britt, 203, 219
Cuyckens, Hubert, 19, 28, 132, 287, Esser, Jürgen, 116, 129
299–300, 327 Evans, Vyvyan, 28, 33, 41, 57, 79–
82, 94, 96–97, 265, 351–352,
d’Andrade, Roy, 70, 76 374–376
Dąbrowska, Ewa, 19, 28, 342, 345,
351 Fabiszak, Małgorzata, 23, 30
Daelemans, Walter, 283, 300, 303 Fauconnier, Gilles, 182, 197, 324,
Dahl, Östen, 307, 309, 313, 325, 328 328, 361–362, 376
Davidse, Kristin, 21, 28, 277, 300 Feldman, Jerome A., 181, 197
Davies, Mark, 35, 95 Fernando, Chitra, 225, 233, 237
De Clerck, Bernard, 274, 299–300 Fieller, Nick, 27, 30, 326
de Saussure, Ferdinand, 5, 12 Fillmore, Charles J., 2, 5, 7, 19, 30–
De Schutter, Georges, 298, 300 31, 45–46, 56, 180, 182, 197–199
De Smedt, Liesbeth, 28 Finkbeiner, Matthew, 116, 131
De Sutter, Gert, 24, 29, 37–38 Firth, John R., 129, 369, 376
Deignan, Alice, 10, 29, 88, 93, 96 Fischer, Kerstin, 26–27, 31, 43–44,
Delorge, Martine, 23–24, 29 47, 50, 53, 57, 179, 186–196,
Deshors, Sandra C., 23, 29, 344, 351 198, 326, 374
Deygers, Katrien, 301 Forsyth, J., 308, 322, 328
Dickey, Stephen, 322–323, 327 Francis, Gil, 129, 369, 377
Diepeveen, Janneke, 329 Fraser, Norman M., 184, 198
Diessel, Holger, 24, 30, 243, 264– Freeman, Margaret, 324, 328
265 Gale, William, 375
Dilthey, Wilhelm, 66, 76 Geeraerts, Dirk, 4–5, 8–11, 14, 16–
Dirven, René, 3–4, 9, 22–23, 26, 29, 17, 19–23, 26–29, 31–32, 35, 37–
31, 35–36, 40, 43, 58–59, 77, 82, 40, 43–48, 51, 54, 57–59, 63, 67,
96, 102, 129, 266, 301, 353, 377 70, 72, 76–78, 80, 94, 96, 99,
Divjak, Dagmar, 11, 15, 21, 23–27, 102–103, 116–117, 129, 132–
30, 34, 47–50, 54–55, 57, 89, 97, 133, 161, 177, 238–240, 242,
239–241, 247–248, 265, 305, 262–268, 271, 273, 275, 286,
313, 326, 328, 333, 341–345, 288–289, 296, 300–301, 303,
347, 351–352, 369, 376 327, 348–349, 353, 357, 367–
Downing, Pamela, 116, 129 368, 373–374, 376
Dowty, David R., 140, 153 Gentner, Dedre, 325, 328
Duinhoven, Anton, 274, 294, 300 Gerrig, Richard J., 363, 376
 Author index 383

Gibbs, Raymond W., 26, 32, 43, 48, Grondelaers, Stefan, 11, 24, 32, 35,
57, 225, 237, 334, 357, 363, 376 39, 45, 57, 116–117, 129, 240–
Gilbert, Nigel G., 184, 198 241, 263–266, 282, 296, 301, 376
Gildea, Daniel, 152 Gumperz, John, 181, 198, 200
Gilquin, Gaëtanelle, 10, 15, 26–27,
32, 126–127, 129, 161, 176, 348, Halliday, Michael A. K, 103, 116,
351 130, 183, 198
Givón, Talmy, 14, 32, 271, 300 Hampe, Beate, 34, 87, 97, 126, 130,
Glucksberg, Samuel, 225, 233, 236– 143, 153, 237, 280–281, 283–
238 284, 301, 348, 352
Glynn, Dylan, 1, 11, 15, 19, 21–30, Han, Chung-Hye, 183, 198
33, 35–37, 39, 50, 54, 57, 90, 96, Hanegreefs, Hilde, 21, 35
239–240, 247–248, 262–266, 268 Hanks, Patrick, 107, 128, 335, 352,
Goddard, Cliff, 366, 376 369, 375, 377
Goldberg, Adele E., 19, 33, 129, Harley, Heidi, 276, 301
155, 161, 172–173, 176, 181, Harris, Roy, 95, 97
198, 204, 219, 226–227, 237, Haspelmath, Martin, 139–140, 145,
271–273, 275–277, 280, 287– 153
289, 292–293, 300 Hay, Jennifer, 141, 153
Goldbeter, Albert, 305, 328 Hebda, Anna, 30
Gonzalez-Marquez, Monica, 35, 40, Herriman, Jennifer, 277, 301
266, 268, 375 Herskovits, Anna, 102, 130
Goodwin, Charles, 82, 96 Heylen, Kris, 10–11, 21, 24, 35, 37–
Goossens, Louis, 4, 29, 33, 38 38, 40, 58–59, 80, 99, 103, 133,
Gordijn, Jaap, 294, 303 240, 264, 266–267, 296, 301,
Gray, Wayne, 132, 200, 379 348–349, 353
Green, Melanie, 79–82, 94, 96, 271, Heyvaert, Liesbet, 155–156, 159,
275, 301, 374–375 175–176
Greenbaum, Sidney, 97, 177 Hilpert, Martin, 20, 24, 26–27, 34–
Gries, Stefan Th., 10–12, 15, 20–21, 35, 47, 49, 52, 155, 161, 176,
23–24, 26–27, 29–30, 32, 34–36, 240, 267, 281, 301
38, 40–41, 43, 47, 49, 52, 54, 57– Hindle, Donald, 375
58, 87, 89, 92–94, 97–98, 107, Hinrichs, Lars, 335, 352
111–112, 114–115, 126–127, Hoey, Michael, 110, 120, 128, 130,
129–130, 133, 143–144, 146, 132, 343, 373, 377
153–154, 156, 161, 163, 176– Hoffmann, Sebastian, 117–118, 120,
177, 230, 235–241, 247–248, 127, 130
265–266, 268–269, 271, 273, Holyoak, Keith J., 324, 328
276, 279–281, 283–284, 296, Huddleston, Rodney D., 86, 97, 278,
301, 303, 333, 335, 339–343, 302
345, 347–348, 350–353, 356, Hunston, Susan, 82, 93, 97, 131,
369, 375–377, 379 369, 377
384 Author index

Israel, Michael, 202, 204, 219 Labov, William, 67, 77, 334, 352,
Itkonen, Esa, 67, 77 367, 378
Lakatos, Imre, 74, 77
Janda, Laura, 11, 19, 23, 36, 126, Lakoff, George, 1–8, 14, 19, 26, 36,
128, 306, 310–311, 323, 328, 45, 58, 102, 131, 180–181, 199,
342–343, 352, 370, 377 241, 244, 262, 267, 294, 302,
Janssen, Theo, 76, 274, 278, 302, 361–362, 378
305, 307, 324, 327 Langacker, Ronald W., 1–7, 13, 19,
Jescheniak, Jörg D., 119, 128, 130 26–27, 37, 45–46, 58, 115–116,
Jing-Schmidt, Zhuo, 23, 36 118, 127, 131, 139, 153, 155–
Johnson, Christopher R., 180, 198 156, 158–159, 161, 166, 175–
Johnson, David M, 200, 379 176, 180–181, 194, 199, 204,
Johnson, Mark, 294, 302 219, 237, 262, 267, 272, 276,
Jones, Steven, 27, 36 280, 302, 357, 361–365, 370, 378
Joos, Martin, 65, 77 Lankenau, Axel, 185, 199
Joseph, Brian D., 95, 97, 128 Lantolf, James P., 95, 99
Lascarides, Alex, 229, 236
Kabata, Kaori, 23, 36 Lee, David, 374, 378
Kay, Paul, 45, 56, 181, 198–199, Lee, Jeong-Hwa, 23, 36
367, 375 Leech, Geoffrey, 97, 162, 177
Keenan, Janice, 366, 377 Lees, Robert B., 156–158, 161, 167–
Keller, Bill, 229, 238, 379 171, 174, 177
Kemmer, Suzanne, 37, 58, 70, 77– Lehrer, Adriene, 3, 31, 37, 56, 198,
78, 143, 152, 219, 237, 301–302 367, 378
Kemper, Susan, 363, 377 Leiss, Elisabeth, 305, 308, 327
Kendall, Tyler, 84–85, 97 Levelt, Willem J. M., 119, 130
Kennedy, Christopher, 141, 153 Levin, Beth, 141, 153
Kennedy, Graeme, 369, 377 Levinson, Stephen, 183, 197, 199
Keppel, Melissa, 237 Levshina, Natalia, 20, 23, 37
Kilgarriff, Adam, 90, 92–93, 97–98, Lewandowska-Tomaszczyk, Barba-
112–113, 115, 126, 129, 131, ra, 21, 23, 26, 29–30, 37, 314,
263, 267 330, 351–352, 377
Kintsch, Walter, 366, 377 Lieber, Rochelle, 218
Klavan, Jane, 24, 36 Lin, Dekang, 229, 237
Knobel, Mark, 116, 131 Lindner, Susan J., 102, 131, 367,
Kokorniak, Iwona, 30 378
Kövecses, Zoltán, 19, 36, 244, 267 Lowe, Will, 369, 378
Krawczak, Karolina, 15, 24, 30, 33,
264 Mair, Christian, 127, 130–131
Krieg-Brückner, Bernd, 199 Majid, Asifa, 37
Krug, Manfred, 118, 126–127, 131 Manning, Jocelyn, 14, 30
Küchenhoff, Helmut, 112, 114, 132 Marzo, Stefania, 26, 37–38
Kuhn, Thomas, 74, 77 Maslov, Jurij S., 310, 326, 329
Matthiessen, Christian M., 131
 Author index 385

Maxwell, Kerry, 121, 131 Panther, Klaus-Uwe, 191, 193, 200,

McCarthy, Diana, 229, 238 272, 302
McCulloch, Charles E., 320, 329 Pedersen, Ted, 153
McDaniel, Chad K., 181, 199 Perkins, Revere D., 312, 327
McEnery, Tony, 82, 98, 344, 353 Peterson, Robert R., 225, 238
McGlone, Matthew S., 225, 238 Petruck, Miriam R. L., 180, 198
Meck, Warren H., 306, 327 Pinker, Steven, 271, 275–277, 302
Medina, José, 325, 328 Popper, Karl, 4, 74, 77, 372–373,
Mervis, Carolyn B., 132, 200, 363, 378
367–379 Porzel, Robert, 197
Meyer, Antje, 119, 130, 302 Prediger, Dale J., 207, 219
Meyer, David, 293 Prevignano, Carlo L., 181, 200
Michaelis, Laura A., 153, 173, 177 Pullum, Geoffrey K., 161, 177, 278,
Miller, George A., 353, 366, 378 302
Miozzo, Michele, 128 Putseys, Yvan, 29
Mittleberg, Irene, 33, 35, 40, 375
Moens, Marc, 140–141, 153 Quirk, Randolph, 97, 156, 177
Mok, Eva, 180, 195, 197, 199
Mukherjee, Joybrato, 27, 37, 104, Rassudova, Ol’ga P., 308, 310, 326,
116, 126, 131 329
Murphy, Lynne, 36 Reppen, Susan, 197
Rice, Sally, 4, 23–24, 26, 34, 36–39,
Narayanan, Srini, 152 58, 88, 97–98, 130, 241, 267,
Nayak, Nandini, 225, 237 334, 352, 379
Newman, John, 12, 14–16, 19, 23– Robinson, Justyna, 21–22, 26, 28–
24, 26, 34, 36–38, 44, 48–49, 51– 30, 33, 35–37, 39, 264, 266, 268
53, 58, 79, 85, 88, 95, 97–98, Röfer, Thomas, 199
126, 130, 180, 199, 262–263, Rosch, Eleanor H., 116, 126, 132,
267, 292, 302, 352 180, 200, 363, 367, 378–379
Newmeyer, Frederick, 172, 177 Rubin, Edgar, 160, 177
Nikitina, Tatiana, 28, 265, 299, 327 Rudzka-Ostyn, Brygida, 4, 9, 19,
Nisbett, Richard, 67, 77 23–24, 29, 37–38, 102, 132
Noël, Dirk, 24, 38 Ruiz de Mendoza Ibáñez, Francisco
Nokkonen, Soili, 23, 38 José, 31, 43, 58, 77, 96, 196, 200,
Nunberg, Geoffrey, 224, 238 377
Nuyts, Jan, 31, 199, 312–313, 317, Rychly, Pavel, 90, 98
329
Sag, Ivan A., 238
O’Connor, Mary, 56, 198 Sampson, Geoffrey, 68, 77
Otani, Naoki, 89, 97, 343, 352 Sander, Keith, 197
Sandra, Dominiek, 4, 29, 31, 39, 77,
Padučeva, Elena V., 305, 308–309, 116, 132, 177, 241, 267, 334,
325, 329 351, 379
Paglucia, William, 312, 327
386 Author index

Sasse, Hans-Jürgen, 138–139, 151, Spivey, Michael, 33, 35, 40, 43, 57,
153 266, 268, 375–376
Sbisá, Marina, 183, 200 Steedman, Marc, 140–141, 153
Schachter, Paul, 156–158, 161, 175, Stefanowitsch, Anatol, 6, 14, 16, 20,
177 24, 26–27, 30, 32, 34, 38–41, 43,
Schegloff, Emanuel A., 181, 200 46–49, 52, 54–58, 92, 94, 97–98,
Scheibman, Joanne, 88, 98 107, 111–112, 114, 127, 129–
Schermer-Vermeer, Ina, 274, 287, 130, 133, 143–144, 153–154,
298, 302 156, 161, 163, 177, 196, 200,
Schmid, Hans-Jörg, 4, 9, 14, 16, 23, 230, 236–238, 240, 242, 265–
27, 33, 39, 44, 47–49, 52–53, 92– 266, 268–269, 271, 273, 276,
93, 96, 98, 101–107, 111–112, 279–281, 283, 301, 303, 347,
114–117, 122, 126–129, 132, 351, 353, 355–357, 369, 376–
247, 266–267, 342, 347, 367, 377, 379
374, 379 Stubbs, Michael, 95, 99, 107, 133,
Schönefeld, Doris, 34, 87, 97, 126, 369, 379
130, 143, 153, 176, 237, 280– Suzuki, Ryota, 340, 353
281, 283–284, 301, 344, 348, 352 Svartvik, Jan, 97, 177
Schulze, Rainer, 4, 39, 102, 132 Swinney, David A., 119, 133
Schütze, Carson T., 68, 77, 334, 353 Szelid, Veronika, 21, 23, 40, 264,
Schvaneveldt, Roger, 293, 302 268
Schwartz, Richard D., 67, 78 Szmrecsanyi, Benedikt, 11, 24, 27,
Scott, Mike, 84, 98 40, 243, 250, 264, 268, 296, 303,
Searle, Shayle R., 82, 320, 329 335, 352
Seechrist, Lee, 67, 78
Sethuraman, Nithya, 280, 300 Talmy, Leonard, 2, 5, 16, 19, 32, 40,
Shimodaira, Hidetoshi, 340, 353 52, 156, 159–160, 167–168, 171,
Simon, Herbert A, 38, 67, 76 177, 240, 268, 281, 300, 303,
Sinclair, John M., 95, 98, 125–126, 337, 353, 365, 379
129–130, 132–133 Tannen, Deborah, 181, 200
Šmelev, Andrej, 305, 308, 310, 329 Tao, Hongyin, 88, 90, 99
Smith, Carlota, 139–140, 142, 145, Taylor, John, 4, 29, 200, 289, 303
154, 316, 329 Teubert, Wolfgang, 80, 99
Smrz, Pavel, 90, 98 Thagard, Paul, 324, 328
Soares da Silva, Augusto, 23, 39 Thornburg, Linda, 191, 193, 200
Solovyev, Valery D., 11, 23, 36, Thorne, Steven L., 95, 99
342, 352, 370, 377 Titone, Debra A., 225, 238
Sonomura, Marion O., 225, 238 Tognini Bonelli, Elena, 104, 129,
Speelman, Dirk, 9, 11, 21, 23–24, 133
27, 29, 32, 35, 37, 39–40, 70, 78, Tomasello, Michael, 26, 40, 152,
161, 177, 236, 238, 240–241, 204, 218
263–264, 266–268, 282, 296, Trnavac, Radoslava, 308, 329
301–302 Tugwell, David, 90, 98
Sperber, Deirdre, 183, 200
 Author index 387

Tummers, José, 10–11, 24, 35, 39, Wasow, Thomas, 238

40, 43, 59, 80, 99, 103, 133, 161, Webb, Eugene J., 67, 78
177, 240, 263, 268, 336, 348– Weijnen, Antonius A., 294, 303
349, 353, 357 Wichmann, Anne, 82, 99
Turner, Mark, 180, 182, 197, 199 Wiechmann, Daniel, 15, 24, 41, 126,
133, 161, 177, 374
Ungerer, Friedrich, 105, 374, 379 Wiemer, Björn, 305, 307–308, 329
Wierzbicka, Anna, 13–14, 19, 41,
van Aken, Hilde, 301 70, 183, 200, 245, 268, 275, 303,
Van Belle, William, 57, 273–274, 365, 379
278–279, 300, 303 Willners, Caroline, 36
Van den Heede, Vicky, 301 Wilson, Dan, 183, 200
Van Langendonck, Willy, 57, 273– Wilson, Timothy, 67, 77
274, 278–279, 300, 303 Wittgenstein, Ludwig, 368, 379
Vendler, Zeno, 140, 152, 154 Wray, Alison, 127, 225, 238
Verhagen, Arie, 70, 77–78 Wulff, Stefanie, 11, 24, 27, 35, 41,
Verkuyl, Henk J., 140, 154 47, 53, 127, 133, 223, 236, 238,
Verspoor, Marjolijn, 82, 96 240, 269, 281, 303, 365, 379
Vorlat, Emma, 29
Xiao, Richard, 82, 98, 344, 353
Wälchli, Bernhard, 313–314, 329
Walter G. Charles, 353 Zaliznjak, Anna A., 310, 329
Warren, Beatrice, 203, 219 Žic Fuchs, Milena, 352, 375
Subject index

aan-Dative Construction, 54, 271– BOTHER, 23

277, 279, 281–300 British National Corpus (BNC), 38,
action-referrent, 155, 157–159, 167, 83–86, 89–91, 95, 107, 112, 145–
169, 171, 174, 175 155, 162–175, 177, 202, 217,
active voice, 24, 67, 247, 347 223–224, 228, 230, 233
adjective, 23, 41, 88, 158, 201–202,
205, 207, 209, 211, 217, 247, Chi-square test, see Pearson’s Chi-
268, 343, 374, 379 square test
adposition, 24 CLAN tool, 84
adverb, 23, 87, 158, 189, 278 coercion, 51–52, 137–138, 140–142,
adverbial, 138–146, 149, 151, 153, 146, 153, 173–174, 177
187–188, 190, 307, 309, 314, cognition, 22–23, 27, 31–33, 39–40,
317, 320, 322 78, 88–89, 103, 116, 128–130,
alternation, 18–19, 23, 25, 34, 57, 132, 180, 184, 199, 237–238,
130, 153, 266, 281, 299–301, 377 267, 302, 305, 328, 357, 376, 378
American National Corpus (ANC), Cognitive Grammar, 33, 37, 58, 131,
83 176, 199, 237, 322, 357, 378
ANGER, 23, 36, 40, 244, 267 Cognitive Linguistics, 2–6, 8–10,
annotation, 23, 82, 240, 244–245, 12–14, 21–22, 25–26, 28–45, 48,
250, 260, 262, 307, 313–314, 51, 55, 57–59, 70, 76–77, 79–83,
336, 347 86, 94, 96–99, 103, 115, 126,
argument structure, 25, 33, 51, 99, 129–130, 132–133, 152–153,
129, 152, 165, 173, 176, 198– 176–178, 180, 195, 197, 199,
199, 237, 243, 245–246, 253, 237, 240–241, 244, 256, 263–
271–272, 275, 277, 279–280, 268, 272, 300–301, 303, 324,
283, 285, 297, 299–300, 302, 352 327–328, 333–335, 337–349,
aspect, 25, 52, 54, 137–138, 140– 351–353, 355–358, 361–365,
142, 145–147, 149, 151–153, 367, 369, 373–380
167, 203, 305–309, 311–318, cognitive processing, 361–364, 373
320–322, 324–326, 328, 344 Cognitive Semantics, 1, 4, 7, 10–11,
atelic, 52, 137–140, 146–151, 170 15, 18–19, 21, 23–25, 28, 30–31,
33–34, 40, 43–55, 57, 61, 79–81,
begin, 39, 102, 132, 311, 339, 342– 96–97, 102–103, 127, 138, 176–
343, 379 177, 180, 199, 217, 225, 234,
behavioural profile, 8, 11, 20, 30, 47, 239, 241, 264–268, 303, 331,
51, 55, 57, 89, 97, 239, 333, 335, 333, 335, 337, 342, 346, 349,
338–346, 349, 351–352, 376 351, 353, 355–356, 359, 361,
bother, 23, 54, 239, 244–246, 252– 365–368, 370–374, 376–377, 379
256, 260
390 Subject index

collexeme, 34, 49, 52, 55, 58, 130, 219, 223–238, 244–245, 247,
143, 146, 148–149, 163, 267, 251–255, 257–261, 263, 265,
285, 290, 295 268–269, 271–273, 275–277,
Collexeme Analysis, 49, 54, 155– 279–281, 283–284, 287, 290,
156, 161–164, 167, 171, 174, 292, 297, 300–301, 303, 307–
271, 273, 275, 279–282, 285– 309, 321–322, 328–329, 342,
286, 296–297 347, 351–352, 365, 379
Collins Birmingham University Construction Grammar, 19, 33, 38–
International Language Database 39, 56, 96, 98, 129, 137, 153,
(COBUILD), 105, 121, 126 155, 176–177, 179–181, 196–
collocation, 8, 10, 22, 25, 88–89, 198, 223, 226–227, 237–238,
129, 172–174, 203, 215, 223, 268, 271–272, 275, 287, 300
236, 240, 369 constructional semantics, 28, 40,
collostruction, 41, 101, 133, 280, 137–138, 144, 147, 151, 161–
283–284 162, 165–166, 171, 223, 271–
Collostructional Analysis, 8, 22, 25, 272, 275, 281, 286, 295, 353
27, 34–35, 52, 57, 85, 92–93, contrastive linguistics, 37–38, 22,–
111, 130, 132, 137, 143–146, 23, 25, 30–31, 38, 55, 236, 300,
148–153, 155, 176, 266, 301– 351
302, 369, 377 conversation, 37, 58, 79, 82–85, 88,
communication, 47, 79, 81–83, 91, 96, 98, 105, 193, 198, 200, 368,
95, 169, 179, 184, 186, 188, 190– 378
191, 194, 197–198, 286, 296 copula, 13, 35, 104, 106, 339
competence, 2, 5–7, 21, 26, 38, 82, corpus, 1, 3–5, 8–22, 26–27, 29–30,
188 32–47, 49–54, 57, 72, 79–80, 82–
compositionality, 47, 51, 53, 132, 88, 90–95, 97–99, 101–125, 129–
223–230, 232–234, 236, 238 130, 132, 139, 143–145, 153,
IDEA, 39, 132, 379 155–156, 159, 161–162, 164,
conceptualisation, 29, 32, 45, 47, 171, 176, 179, 184, 189, 195–
182, 184, 188, 306, 328, 361– 196, 201–202, 205, 217–219,
364, 367, 373 223–224, 228–230, 233–234,
Configural Frequency Analysis, 22– 236–237, 239–240, 242, 244–
24, 351 246, 262–267, 271–272, 274,
confirmatory statistics, 21, 23, 50, 276, 279–280, 282–286, 293–
243, 318 294, 296–299, 301–302, 305,
constructicon, 223, 226–227 307–311, 313–315, 317, 321–
construction, 18–19, 23, 25, 29, 33– 322, 325, 328, 333–338, 340,
35, 37–41, 47, 49, 52–53, 56, 58, 342–343, 345–352, 355, 357,
78, 83, 92, 96–98, 101, 104, 106– 364, 367, 368–370, 373–374,
113, 118, 120, 122–124, 126, 376–377, 379
129–133, 137–141, 143–146, corpus linguistics, 10, 15–16, 43, 80,
149–151, 153–155, 160–163, 91–92, 120, 143, 244, 333, 335–
167, 176–177, 179–181, 183, 337, 350, 364
190–191, 195–199, 201, 203,
 Subject index 391

Corpus of Contemporary Written 156, 161, 163–166, 172, 176–

Dutch (ConDiv), 282, 298, 301 177, 184, 200, 203, 218, 223,
Corpus Search, Management and 228, 238–239, 265, 267–268,
Analysis System (COSMAS), 205 271–278, 280–283, 285, 287–
Correlation Analysis, 50, 201, 208– 288, 292–293, 296, 298–303,
209, 211, 213, 215 319, 333, 335, 342–344, 346–
Co-varying collexemes, 49, 58, 236– 347, 350–352, 367, 377–379
237, 301 entrenchment, 3, 5–6, 8, 13–14, 27,
C-value (C-score, C-index), 26, 64, 39, 47–48, 52, 98, 101–103, 115–
92, 99, 112, 146, 150, 200, 230– 128, 132, 218, 223, 228
232, 237, 258–259, 279–280, Estonian, 24, 36
321–322, 326, 351 exploratory statistics, 22–23
Czech, 36, 292, 322
factive, 155, 157–159, 168, 170–
dative, 24, 28, 31, 38, 54, 57, 265, 171, 174–175, 305
287, 299–300, 302–303 features, 11, 17, 20–22, 25, 32, 44,
declarative, 183, 187–190, 194, 200, 47, 52, 54, 64, 69, 71–72, 75, 84–
309, 339 85, 92, 142, 179–180, 182, 184,
diachronic, 22–23, 25, 29, 32, 34– 188,–190, 196, 239–245, 247–
35, 38, 129, 266, 290, 294, 301, 256, 259–261, 273, 275, 308,
368, 376 339, 366–367, 370
Distinctive Collexeme Analysis, 267 Finnish, 23, 342, 350
Ditransitive Construction, 24, 37, Fisher-Yates Exact test, 111–114,
131, 226 122–124, 127, 144, 162–163,
Double Object Construction, 54, 230, 258, 280, 283
271–277, 279–281, 286–288, for, 140
290–293, 296–297, 299, 301–302 force dynamics, 52, 155–156, 161,
durative, 52, 137–140, 142, 144– 165, 167, 171–173, 176, 281, 302
152, 248, 316 form-meaning pair, 81, 143, 181,
Durative Adverbial Construction, 226, 253, 256
137–140, 142, 144–152 Frame Semantics, 56, 180, 197, 245
Dutch, 23, 28–29, 31, 35, 37–40, 54, French, 23, 29, 128, 344, 351
57, 70, 78, 176–177, 236, 238, frequency, 6, 8, 14, 17, 26–27, 39,
266–268, 271–278, 281–288, 47, 49, 52–53, 85, 88, 90, 92–93,
290, 292–294, 296–303, 329, 367 98, 101–102, 104, 106–112, 115–
131, 133, 137, 139, 143–144,
Embodied Construction Grammar, 162–163, 188–189, 201–203,
179, 181–182, 192–193, 195 205, 209, 211, 216–218, 227–
empirical cycle, 17, 63, 73–75, 262 229, 231, 233, 284–285, 334,
encyclopaedic semantics, 6–7 340, 345–346, 352, 359, 367
English, 13, 23, 28-30, 35, 37–41,
52–54, 84–85, 87, 91, 95–98,
102, 105, 128–133, 137–140,
145, 147, 149–151, 153, 155–
392 Subject index

genitive, 24 introspection, 2, 4, 16, 17, 21, 63,

German, 23, 35, 201–202, 205, 207, 66–68, 71, 240, 243, 250, 286,
212, 266–267, 272, 281, 292, 309, 335–337, 374
296, 298, 301, 344 intuition, 4, 54, 63–64, 66–67, 70–
gerund complement clause, 159, 71, 75, 89, 111, 123, 130, 214,
162–165, 168, 172–174, 281 223, 308, 336
going-to Future Construction, 347
grammaticality, 2–4, 6, 8, 13, 68, 77, Japenese, 23
337
Kappa (Prediger’s free-marginal or
HAPPY, 23, 36 Cohen’s), 207, 210, 215
HATE, 23, 30, 36 Korean, 23, 36
hermeneutics, 63, 65, 240
Hierarchical Cluster Analysis Lancaster-Oslo/Bergen Corpus
(HCA), 22–24, 27, 49, 257, 338, (LOB), 347
340, 342 langue, 2, 5–8, 12
Hungarian, 23 lemma, 19, 52, 79, 87, 89–91, 95,
262, 280, 295, 336, 338, 347
iconicity, 30, 265, 268 lexeme, 9, 92, 102, 111–112, 119,
ideal speaker, 2, 4, 5, 66–67 138, 141, 143–144, 163, 244
ID-tags, 55, 338–343, 345, 350, 375 lexical semantics, 4, 18, 25, 27, 39,
imperative, 183–184, 187–191, 194, 51, 54, 65, 99, 132, 158–159,
200, 312, 339 180, 241, 345, 349
imperfective, 23, 50, 139, 248, 305– lexicon, 30–31, 56, 67, 116–117,
311, 315–319, 321–326 132, 138, 198, 203, 219, 225–
infinitive, 52, 87, 92, 104, 106, 110, 227, 234, 256, 297, 302, 307,
113, 121–122, 124, 155–157, 365, 375, 378
161–162, 164–165, 171–172, LIBERTY, 23, 33
175, 281, 307–309, 314–322, 326 Linear Discriminant Analysis
inflection, 24–25, 39, 40, 88, 177, (LDA), 21–22, 24, 344
205, 236, 238, 268, 343 linguistic evidence, 35, 63, 266, 301
International Corpus of English Logistic Regression Analysis (LRA),
(ICE), 83, 87, 145–146, 150, 340, 21–22, 27, 39, 48, 50, 239, 243,
347 257–259, 264, 268, 296, 307,
interpretation, 11, 28, 44–45, 63–65, 318–321, 325–326, 344
67, 69–70, 73, 75, 111, 146, 148, LOVE, 23, 30, 214, 244, 337, 267
173–175, 180, 187, 198, 209,
229, 237–238, 257, 265, 274, Mandarin, 23, 36, 344
286, 299, 317, 326–327, 337, 375 metaphor, 7, 19, 33, 45, 65, 199,
interrogative, 183, 187–188, 190, 210, 212–214, 227, 264, 289,
339 294, 323, 328, 363, 377
into-Causative Construction, 24 methodology, 2, 10, 27, 43–44, 47–
48, 63, 65, 68, 71–72, 76–77,
 Subject index 393

103, 156, 161, 163, 264–265, operationalisation, 3, 6–8, 13, 15–16,

268, 333, 352–353, 355, 357, 374 44, 46, 55, 73, 117, 208, 218,
metonymy, 7, 33, 65, 212–213, 324, 240, 242, 244, 249, 309, 338,
363 355–356, 358–361, 363–364,
Mixed Effects Logistic Regression 366–368, 370, 372–373
(MER), 22, 27, 305, 307, 320,
321, 326 Parallel Corpus of Slavic and Other
modality, 51, 54, 155–156, 175, 305, Languages (ParaSol), 314, 330
307–309, 312–313, 315–329, 351 parts of speech, 81, 85, 87, 91, 245
morphology, 7, 19, 141, 203, 219, part-of-speech tags (POS), 79, 85,
241, 298 283
multicollinearity, 258–259 passive voice, 24, 49, 67, 90, 300,
multidimensional approach, 271, 341, 347
273, 286 Pearson’s Chi-square test, 21, 94,
Multidimensional Scaling (MDS), 112, 259, 345
22–23, 28 Pearson’s R correlation, 211, 213,
multifactorial, 8–9, 17, 22, 29, 33– 215
34, 36, 41, 54, 240–241, 247, penalised maximum likelihood, 258
257, 266, 273, 288, 296, 342, 379 perfective, 23, 50, 139, 248, 305–
Multiple Correspondence Analysis 326
(MCA), 21–23, 27, 239, 251–252, perspective, 26, 28, 34, 37, 40, 44–
254–255, 257, 260, 264 46, 52, 54, 57, 66–67, 73, 97,
multivariate statistics, 9, 10, 17, 21, 108, 118, 130, 151, 153, 156,
29, 125, 239, 241, 244, 263, 296, 163, 176, 180, 185, 224, 226–
352 227, 230, 240, 242, 266, 268,
271, 277, 281, 294, 296, 299,
načat'sja, 343 301–303, 305, 345, 350, 352–
načat’, 343 353, 356, 360–361, 374, 377
Natural Semantic Metalanguage phasal verbs, 30, 342–344, 351
(NSM), 70, 365–366 philosophy of science, 355, 358,
need to, 24 371, 375
nominalisation, 155–156, 176–177 Polish, 23, 28, 30, 38, 314, 322, 328,
norovit’, 341–342 330
noun, 23, 25, 39, 88–90, 93, 104, polysemy, 4, 18–19, 29, 31, 33, 54–
106–111, 113–116, 118–120, 55, 57, 102, 172, 239–241, 261,
122–124, 129, 132, 163, 180, 265–266, 280, 333, 336, 339–
189, 208, 211, 236, 244–245, 340, 345
247, 267, 367 Portugese, 23
preposition, 24–25, 28, 37, 91, 117,
objectivity, 20, 22, 54, 64, 73, 75, 130, 140, 145, 160–161, 164,
239–240, 242, 262, 316, 345, 350 244, 272, 276, 278, 282, 287,
onomasiology, 18–20, 22, 25, 31–33, 292, 298–300, 303
116–117, 239, 261–263 Principal Component Analysis
(PCA), 22, 24, 347
394 Subject index

probability value (p-value), 11, 111– semantic weight, 223

115, 120–124, 146, 163, 235, semantics, 1–2, 5–11, 16, 18, 20–21,
258–259, 280, 283, 340 25, 30–31, 33, 36, 45–47, 51–52,
probovat’, 341 54–55, 63–65, 68, 71, 76, 79–81,
productivity, 14, 48, 201, 203, 209, 102, 129, 137–139, 142–143,
211, 217, 219, 238 145, 151, 153, 159, 180, 196,
Profile-Based Analysis (PBA), 22– 200, 225, 227, 230–231, 234,
23 236, 239–240, 242, 244, 247,
progressive, 113, 148, 153, 169 250, 256–257, 263–264, 266–
prominence, 45, 88, 227 267, 271, 273–275, 280, 286–
pronouns, 110, 157–158, 298 288, 296, 300–301, 303, 305,
prosody, 7, 19, 25, 56, 241, 263, 353 311, 322, 324, 333–334, 337,
proyvat’sja, 341–342 346, 355–356, 358, 361, 363,
pytat’sja, 341 365, 368, 371–376, 379
pyzit’sja, 342 semasiology, 18–19, 22, 25, 239,
260–263
quantitative, 21, 44, 47–55, 72, 79, semelfactive, 139–140, 142, 311
93, 101, 103, 123, 125, 180, 195, sentence mood, 179, 183, 189, 191,
223, 234, 241–242, 333, 344, 194–195
348, 377 sequential scanning, 141
shell nouns, 104, 106, 110, 113, 116
R, 26, 28, 77, 127, 144, 153, 163, silit’sja, 341–342
176, 198, 200, 208, 230–236, Slovak, 322
238, 258, 265, 283, 301, 321, Slovene, 322
327, 351, 353, 376 sociolinguistics, 1, 9, 11, 13, 21–23,
raising, 24 25, 28, 31–33, 35–36, 38–40, 58,
Receiver Operating Characteristic 67, 97, 181, 186, 241, 251, 263,
curve (ROC), 321 266–268, 301, 303, 353, 373
R-squared value (R2-value), 230- Somer’s Dxy rank correlation, 258,
236, 258–259 322, 326
Russian, 23, 30, 36, 50, 54, 305–311, Sorbian, 322
313–314, 316–323, 325, 327– SPSS, 127
330, 333, 341–345, 351–352, 377 starat’sja, 341
start, 39, 64, 90, 102, 132, 172, 204,
SAD, 23 276, 309, 311, 339, 342–343, 379
salience, 14, 27, 30–31, 102–103, stat’, 311, 343
116–117, 127, 132, 295, 323 statistical significance, 11, 17, 21,
semantic extension, 48, 201–202, 92, 113, 190, 243, 251, 259, 263,
204, 288, 292, 297 350
semantic frame, 19, 166, 168–169, Student's t-test (t-test), 21
171 Swedish, 23, 281
semantic role, 303, 338, 343 synonymy, 18–19, 27, 30, 33, 36, 39,
semantic similarity, 54, 213, 223, 41, 55, 58, 97, 239–241, 268,
229, 231, 353
 Subject index 395

303, 333, 339, 341–343, 345, 238–240, 265, 268, 271–272,

350–353, 377 280, 301–302, 305, 333–334,
syntax, 2, 7, 9, 12, 19, 24, 30, 32–34, 348–350, 352, 377, 379
38, 40–41, 57, 96–97, 129, 153, usage-based model, 13, 58, 152
177, 203, 225–226, 241, 245,
256, 263, 265–266, 268, 275, V NP-Construction, 54, 223, 227,
303, 333, 351, 365, 376 231, 235
Verb-Particle Construction, 149, 347
telic, 52, 138, 141, 146–149, 151, verbs, 23–25, 28–29, 34, 36–38, 52,
316 86, 88, 90, 92, 97, 105, 113, 126,
temporal adverbial, 30, 142, 265 130–131, 137, 139–140, 145,
tief, 201–202, 205–214, 217 149– 153, 159–160, 163, 165,
transitive, 30, 87, 165, 168, 175, 168–174, 180, 187, 194, 219,
247, 339 227, 234, 236–238, 243–247,
tscit’sja, 342 268, 271, 273–275, 278, 280,
tuzit’sja, 342 282– 286, 289– 293, 295, 297,
t-values, 341 299, 301, 306, 311, 316, 328,
type frequency, 209 341– 345, 352, 366, 369

Ukrainian, 23, 322 will Future Construction, 347

usage-based, 1, 3, 5–6, 10, 12–13, Wordsmith Tools, 84, 98
28, 33, 35, 37, 39–41, 43, 46, 52,
58, 80–82, 94, 96, 99, 131, 137, z-value (z-score, standard score), 85,
139, 143, 155, 176, 179–181, 93, 230, 258–259, 342
201, 203–204, 217–219, 223,