Creativity and Artificial Intelligence

≥
Applications of Cognitive Linguistics
4

Editors
Gitte Kristiansen
Michel Achard
René Dirven
Francisco J. Ruiz de Mendoza Ibáñez

Mouton de Gruyter
Berlin · New York
 Creativity and
Artificial Intelligence
A Conceptual Blending Approach

by
Francisco Câmara Pereira

Mouton de Gruyter
Berlin · New York
 Mouton de Gruyter (formerly Mouton, The Hague)
is a Division of Walter de Gruyter GmbH & Co. KG, Berlin

앪
앝 Printed on acid-free paper
which falls within
the guidelines of the ANSI
to ensure permanence and durability.

Library of Congress Cataloging-in-Publication Data

Pereira, Francisco Câmara, 1972⫺

Creativity and artificial intelligence : a conceptual blending ap-
proach / by Francisco Câmara Pereira.
p. cm. ⫺ (Applications of cognitive linguistics ; 4)
Includes bibliographical references and index.
ISBN 978-3-11-018609-3 (hardcover : alk. paper)
1. Creativity (Linguistics) 2. Artificial intelligence. I. Title.
P37.5.C74P47 2007
4011.41⫺dc22
2007035014

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.

ISBN 978-3-11-018609-3
ISSN 1861-4078

쑔 Copyright 2007 by Walter de Gruyter GmbH & Co. KG, D-10785 Berlin
All rights reserved, including those of translation into foreign languages. No part of this book
may be reproduced or transmitted in any form or by any means, electronic or mechanical,
including photocopy, recording, or any information storage and retrieval system, without
permission in writing from the publisher.
Printed in Germany
Acknowledgements

In parallel to the work behind this book, a whole novel could have been writ-
ten, with many more actors than appears in the final document. The truth is
that, without these actors, it would certainly lose much of its value.
My former supervisor and colleague, Prof. Dr. Amı́lcar Cardoso, has been
a key actor from the first to the last chapter, sometimes as a patient advisor
and listener, other times as a friend.
Some colleagues (and friends) had the leading role on occasion, even if
they were unaware of it. In fact, a substantial part of this book was a result
of talks, collaborations and constructive arguments with Tony Veale, Pablo
Gervás, João Leite and Geraint Wiggins.
Also, without the rest of the AILab crew (before and after ECOS), a lot
of science and friendship chapters would certainly be irreparably lost. These
were Penousal, Paulo, Grilo, Xico, Leonor, Prof. Carlos Bento, Prof. Ernesto
Costa, Jorge Tavares, André, Miguel Ferrand and Ana Alves.
Special thanks must be made to Paulo Ribeiro, João Pedro Martins, Bruno
Leitão and Bruno Marques who have contributed with work that is also stron-
gly linked to the thesis that originated the book.
A large part of this novel was done within the environs of the Departament
de Informatics Engineering of the University of Coimbra, and I must not
forget to thank my colleagues there and particularly the staff, Isabel, Paula,
Manuela and Márcia, for their support.
The other side of the story, with different scenery and people, starts at
home. Ana has given the strongest support and love as only she can give and
has frequently drawn from a limitless reservoir of patience. Kali, Becas and
Sansao have also been patient and supportive, in the way that only cats and
dogs can be. Kali and Becas have left us during this period, and we feel a deep
sorrow, for this and future novels would never be the same without them.
Finally, my parents and sister deserve a special mention for the many times
that I missed them, especially when they needed me. And for the support they
have given from the beginning.
I must also kindly thank the reviewers of this (and previous) publicati-
ons for the several constructive and incisive comments that definitely helped
improve its potential as a contribution to science.
Contents

1 Introduction 1
1 Motivations . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
2 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
3 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 4
4 Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Creativity 7
1 Creativity theories . . . . . . . . . . . . . . . . . . . . . . . . 7
1.1 Divergent Production . . . . . . . . . . . . . . . . . . . . . . 8
1.2 Bisociation . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.2.1 Humor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2.2 Science . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.2.3 Arts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.3 The Systems Model . . . . . . . . . . . . . . . . . . . . . . . 19
1.4 Boden’s taxonomies . . . . . . . . . . . . . . . . . . . . . . . 23
1.5 Others . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.6 Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2 Computational Creativity . . . . . . . . . . . . . . . . . . . . 30
2.1 Two accounts for characterizing Creativity . . . . . . . . . . . 30
2.1.1 Characterizing Creativity in AI . . . . . . . . . . . . . . . . . 31
2.1.2 Creativity assessment . . . . . . . . . . . . . . . . . . . . . . 32
2.2 Creative systems . . . . . . . . . . . . . . . . . . . . . . . . 36
2.2.1 What is a creative system? . . . . . . . . . . . . . . . . . . . 36
2.2.2 Some systems and models . . . . . . . . . . . . . . . . . . . 38

3 Working with Concepts 47

1 What is a concept? . . . . . . . . . . . . . . . . . . . . . . . 47
2 Building concepts . . . . . . . . . . . . . . . . . . . . . . . . 49
2.1 Concept Formation . . . . . . . . . . . . . . . . . . . . . . . 49
2.2 Concept Invention . . . . . . . . . . . . . . . . . . . . . . . . 50
3 Mixing concepts . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.1 Conceptual Combination . . . . . . . . . . . . . . . . . . . . 51
3.2 Conceptual Blending . . . . . . . . . . . . . . . . . . . . . . 53
viii Contents

3.2.1 The Framework . . . . . . . . . . . . . . . . . . . . . . . . . 55

3.2.2 Some criticisms . . . . . . . . . . . . . . . . . . . . . . . . . 66
4 Metaphor and Analogy . . . . . . . . . . . . . . . . . . . . . 68
4.1 Structure Mapping Engine . . . . . . . . . . . . . . . . . . . 69
4.2 Conceptual Scaffolding and Sapper . . . . . . . . . . . . . . . 71
4.3 Others . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

4 A Model of Concept Invention 85

1 A Creative General Problem Solver . . . . . . . . . . . . . . . 85
2 Description of the model . . . . . . . . . . . . . . . . . . . . 89
3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5 Divago 101
1 Overview of the architecture . . . . . . . . . . . . . . . . . . 101
2 Knowledge Base . . . . . . . . . . . . . . . . . . . . . . . . 103
3 Mapper . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4 Blender . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
5 Factory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6 Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.1 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.2 Topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
6.3 Pattern Completion . . . . . . . . . . . . . . . . . . . . . . . 131
6.4 Maximization/Intensification of Vital Relations . . . . . . . . 132
6.5 Unpacking . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
6.6 Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
6.7 Relevance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7 Elaboration . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
8 Divago as a creative system . . . . . . . . . . . . . . . . . . . 140

6 Experiments 141
1 The Boat-House . . . . . . . . . . . . . . . . . . . . . . . . . 142
2 The Horse-Bird . . . . . . . . . . . . . . . . . . . . . . . . . 149
2.1 Evaluating the Optimality Pressures . . . . . . . . . . . . . . 151
2.2 Finding the Pegasus . . . . . . . . . . . . . . . . . . . . . . . 154
3 Noun-Noun combinations . . . . . . . . . . . . . . . . . . . . 161
3.1 Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
3.2 Free generation . . . . . . . . . . . . . . . . . . . . . . . . . 166
3.3 Comparison to C 3 . . . . . . . . . . . . . . . . . . . . . . . . 171
 Contents ix

4 The creature generation experiment . . . . . . . . . . . . . . . 173

4.1 Tuning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
4.2 Free generation . . . . . . . . . . . . . . . . . . . . . . . . . 178
5 The established Blending examples . . . . . . . . . . . . . . . 184
5.1 Experiments with isolated principles . . . . . . . . . . . . . . 187
5.2 Experiments with combination of principles . . . . . . . . . . 194
5.3 Some considerations . . . . . . . . . . . . . . . . . . . . . . 196
6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

7 Conclusions and Future Directions 205

A The effect of input knowledge 211

B Established examples of Conceptual Blending 213

C Programming the Frames 223

1 Syntax and Overview . . . . . . . . . . . . . . . . . . . . . . 223
2 Programming of the frames . . . . . . . . . . . . . . . . . . . 223
3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225

D Instances in the house-boat experiment 231

E Experiments, Databases and other documents 233

Notes 235

Bibliography 239

Index 250
Chapter 1
Introduction

It is obvious that invention or discovery, be it in mathematics or anywhere

else, takes place by combining ideas... (Hadamard)

Pour inventer il faut penser à côté (Souriau)

The useful combinations [of ideas] are precisely the most beautiful.
(Poincaré)

1. Motivations

Right from the start, the main focus in AI research has always been with
the issue of problem solving. Seen from this point of view, intelligence cor-
responds to the ability to solve (complex) problems, from the accurate au-
tonomous movement of a robot arm to the understanding of a natural lan-
guage sentence. The classical setting is that of a search in a space of solu-
tions for the problem, where an intelligent agent looks for the best choices.
A commonly used analogy is of this agent travelling in a search space with
mountains (highly valued solutions), valleys (bad solutions), islands (moun-
tains surrounded by valleys) and plains (areas where quality of solution hardly
changes)1 .
One of the most common criticisms made of Artificial Intelligence meth-
ods of problem solving is their limited ability to deal with situations not pre-
dicted in the specification. The search space is normally strictly defined, how-
ever flexible, complex and adaptable the system seems to be. When facing a
problem with no satisfactory solution in its search space, an AI system simply
returns, at best, the least unsuccessful result that exists in that search space-
even when the solution is achievable via the simplest operation of changing
perspective, relaxing a constraint or adding a new symbol. In other words,
such systems are hardly capable of performing what we normally call cre-
ative behavior, a fundamental aspect of intelligence.
2 Introduction

However, the recognition that there is a deficit of creativity within AI sys-

tems does not by itself bring new solutions any more than it reasserts that
computers, as we know them, are formal machines that are limited to their
closed worlds. The question therefore arises about what can be done to make
them more creative or even if, with the current computational architectures,
that is possible at all. To some extent, some of the current state-of-the-art
paradigms (such as Evolutionary Computation, Multi-agent Systems or Case
Based Reasoning) have been responsible for many of the developments re-
garding the first part of the question. Indeed, we have been developing less
rigid systems in previous years and sometimes even producing striking re-
sults. Nevertheless, when any of these systems finds a situation for which it
was a priori not specified to solve, it is definitely not able to cope with it.
The second half of the question concerns primarily what the essential com-
ponents of a creativity model could be and whether these can be present in
a formal machine. There is no definitive answer for this, yet we can allow
ourselves cross-fertilization from other areas, such as Psychology, Cognitive
Linguistics, Cognitive Science and Philosophy, in the speculation and build-
ing of a possible solution.
The relationship between Intelligence and Creativity poses further ques-
tions. Are these two independent properties of cognition or, on the contrary,
are they interrelated and inseparable? More specifically, if taking a traditional
AI perspective: isn’t creativity about search? Is it a different approach to in-
telligence?
These questions are present throughout this book, which is an attempt to
approach them according to a perspective that, while centered on Computer
Science and AI, also lifts contributions from other areas.
The reader can find most of the material described and used in the context
of this book in the attached CD and at http://eden.dei.uc.pt/~camara/
AICreativity. In this website, the reader can find the latest additions to this
and related works.

2. Overview

There is general agreement that the ability to find relations between appar-
ently unrelated knowledge is a creative activity. As can be found in many
studies from the area of cognitive psychology, the creative faculties of the
human mind are hugely dependent on the ability to search through spaces
 Overview 3

or “viewpoints” that are different from the ones immediately or more obvi-
ously involved. For example, according to (Marin and De La Torre 1991),
our capacities of abstraction, symbolic analysis, of finding less-obvious re-
lations, among others, are associated to creative production. Indeed, many
important discoveries, great music compositions or paintings were reportedly
achieved at times of wandering in domains not directly related to the actual
problem (e.g. the famous dream of Kekulé, the discoverer of the structure of
the Benzene molecule, who was dozing by the fire and dreaming of self-biting
snakes when he made his major discovery (Boden 1990)). One of these psy-
chology theories (Guilford 1967) concentrates on the idea of divergent think-
ing. Arthur Koestler (Koestler 1964) also wrote about a related phenomenon,
naming it bisociation. From the computer science point of view, the mod-
elling of divergent thinking and bisociation seems extremely difficult mainly
because it is far from formally specified and, even if it was, it would certainly
demand cognitive capacities that are still not achievable by computers. Yet,
this does not mean that it is impossible to build models, perhaps less ambi-
tious ones, that are capable of achieving a smaller degree of divergence, in
which a computer is able to reason in a multi-domain knowledge base, even-
tually solving problems via transferring knowledge from different domains.
Since different domains will contain different knowledge and possibly differ-
ent symbols and representations, a model for reasoning in a multi-domain en-
vironment must have translation mechanisms, so that the transferred knowl-
edge will still have meaning in the new context. There are well known cogni-
tive mechanisms that establish cross-domain relationships, namely Metaphor
and Analogy, which have been studied to some depth within AI, and which
are certainly good candidates for plausible cross-domain transfer.
A perfect cross-domain transfer mechanism will be futile if the new
knowledge is not integrated into its novel context in a meaningful way. This
integration demands processes and principles able to generate knowledge
structures that can be considered as a whole rather than the sum of its parts. In
other words, the transfer of new knowledge should not be condemned to re-
sult in a pastiche or a concatenation of the parts, instead an emergence of new
structure, function or behavior is to be favoured. Two research trends from
Cognitive Science aim to solve this problem, namely Conceptual Combina-
tion and Conceptual Blending (also known as Conceptual Integration). The
former traditionally deals with the understanding of linguistic combinations
(such as “pet fish” or “mountain stream”) while the latter is conceived as a
process that can apply across the cognitive spectrum in general. Despite their
4 Introduction

differences, they both share the intent of understanding the cognitive abil-
ity of integrating knowledge from distinct sources. Both have already been
subject to computational modelling.
Finally, the unavoidable question of evaluation could justify a research
programme on its own, with worries regarding expertise, intentionality, com-
plexity, aesthetic judgement, constraint satisfaction and novelty, to name only
a few topics. In the current context, the evaluation should be primarily con-
cerned with whether the just created knowledge structures are worth consid-
ering for further use and treatment within the domain it was designed for. In
other words, if it is both novel and useful within this domain. The computa-
tional approach to novelty assessment has been based on similarity metrics
or clustering techniques while determining usefulness is normally done via
application of rules or checking constraint satisfaction. Conceptual Blend-
ing proposes a set of generic Optimality Constraints that aim to govern the
generation of a blend. However, these are not explained formally, raising the
challenge of their computational modelling.
We have just summarized some of the components for a Model of Con-
cept Invention from cross-domain transfer. By concept invention, we mean
the generation and addition of a new concept (or knowledge structure) to
a domain in which this new concept could not be obtained by any internal
means (e.g. deduction) and which can still be accepted as a valid concept for
the domain. For example, before the invention of the airplane, the domain of
“transport means” did not have the entire knowledge to lead to it. It was nec-
essary to observe physical laws that were not taken into account for any other
previous means of transport (even the balloon) in order to create different
concepts of propulsion and sustaining structure.

3. Contributions

The main expected contributions of this book are:

– A reflection, overview and state-of-the-art survey about creativity research,
according to different perspectives such as Philosophy, Psychology, Cog-
nitive Science and Computer Science;
– A formally specified Model of Concept Invention, based on processes and
principles that are coherent with the current research on creativity;
– An implemented system, Divago, which partially instantiates the Model of
Concept Invention. Divago was applied to different domains and demon-
 Structure 5

strated to be capable of generating results that pass the criteria of creativity

assessment used.
– A Computational Model of Conceptual Blending, which will become in-
tegrated within Divago. This is the first computational approach to Con-
ceptual Blending (Fauconnier and Turner 1998) that includes all the fun-
damental aspects of this framework.
– An assessment of the creativity of the results and of the system. We analyze
the creativity of Divago with the frameworks suggested by Ritchie (Ritchie
2001), Wiggins (Wiggins 2001, 2003) and Colton et al (Colton, Pease and
Ritchie 2001).

4. Structure

The remainder of this book is structured as follows:

– Chapter 2 is about research on creativity. It provides the necessary back-
ground regarding theories of creativity, computational approaches to cre-
ativity and frameworks for creativity assessment. In this chapter, the reader
will also find the generic guidelines that underlie the rest of the book,
namely at the level of the Model of Concept Invention (chapter 4) and
of assessment of the results of Divago (chapter 6).
– Chapter 3 starts by defining what a concept is in the context of our work.
It also defines concept invention and compares it with concept formation,
two kinds of concept building processes. Working with concepts is funda-
mental for this book, and specifically the framework of Conceptual Blend-
ing, which is also presented in this chapter. Conceptual Combination, a
related area, is also presented, with particular focus to C3 , a system that
will later on (in chapter 6) be compared to Divago. The chapter ends with
an overview of computational approaches to Metaphor and Analogy, which
deal with concept networks and from which we developed a part of Divago
(the Mapper). After this chapter, the reader will have obtained a first insight
on the practical aspects involved in this work (in chapter 5) and a clearer
impression of the necessary notions regarding concepts.
– Chapter 4 is dedicated to the description and formalization of the Model
of Concept Invention. There, the reader will find a theoretical model, in the
sense that it has not been totally implemented or specified up to the detail of
computational implementation. This model provides a set of modules that
we argue should be present in a system that is meant to invent concepts.
6 Introduction

– Chapter 5 describes Divago in detail, a system that partially implements

the model presented in chapter 4. This description will take into account
the modules of that model (with redefinition of the formalizations when
necessary) and the framework of Conceptual Blending, which is the basis
for the bisociation mechanism of Divago. After finishing this chapter, the
reader will know Divago in depth, namely its knowledge representation,
search mechanism and blending model.
– Chapter 6 is dedicated to the experiments made with Divago. We show
the five different experiments made: house-boat, for analysis of the search
space; horse-bird, for the study of behavior of Divago with regard to the
Optimality Constraints; noun-noun, for the generation of noun noun com-
binations and comparison to C3 ; creatures generation, for the testing of
Divago as an engine for generating concepts in a game environment; and
established blending examples, for the validation of the Blending model
implemented in Divago. The reader will get an idea of the behavior of
the system within these different situations, namely with attention to the
novelty and usefulness of the results. Throughout this chapter, we will ana-
lyze the system according to the frameworks of Ritchie (Ritchie 2001) and
Colton et al (Colton, Pease and Ritchie 2001).
– Chapter 7 is dedicated to the final conclusions and discussion of future
directions to take. There, the interested reader will find a multitude of re-
search directions, some related to the generic aspects of creativity, concept
invention and Blending, some more specifically directed towards the future
developments of Divago.
Chapter 2
Creativity

In the first half of this chapter, we present approaches to creativity within

Psychology, Philosophy, Cognitive Science and AI. The second half is specif-
ically dedicated to the area of computational creativity, where we will show
the state-of-the-art both at the level of the theoretical foundations and at the
level of implementations.

1. Creativity theories

Creativity has been the motivation for many lines of writing throughout hu-
man history, for it is such an appealing and mysterious aspect of our ex-
istence. However, it is also noticeable that its study, from a scientific per-
spective, has been neglected until the second half of the twentieth century
(Albert and Runco 1999). The early twentieth century scientific schools of
psychology, such as structuralism, functionalism and behaviorism, devoted
practically no resources at all to the study of creativity (Sternberg and Lubart
1999). The oft cited foremost turning point was when Joy Paul Guilford,
in his American Psychological Association presidential address, challenged
psychologists to pay attention to what he found to be a neglected but ex-
tremely important attribute, namely, creativity (Guilford 1950). The so called
first golden age of creativity then took place, with many newly founded re-
search institutions. However this revolution did not last for long. In fact, from
1975 to 1994, only about 0.5% of the articles indexed in Psychological Ab-
stracts concerned creativity (Sternberg and Lubart 1999). Today, it seems that
the subject has gained another burst of output (the second golden age). In-
deed, unprecedented resources are being directed towards creativity research
in many areas.
In the following sections, the reader will be introduced to some of the
works on creativity that influenced this book. These works come from the
areas of Psychology (section 1.1), Philosophy (section 1.2) and Cognitive
Science (1.4). Without having a direct influence on our own work, the con-
tribution of Csikszentmihalyi (Csikszentmihalyi, 1996) is also presented in
section 1.3 for three special reasons: it is often cited in works of Creativity
8 Creativity

and AI2 ; it is also a respected and referential work within the area of Psy-
chology; it reasserts some of the conclusions given by the previous sections,
attesting their current acceptance. In section 1.5, we will complete the state-
of-the-art of creativity with an overview of other works. Finally, a synthesis
will be made in section 1.6, with particular emphasis on the aspects relevant
to this book.

1.1. Divergent Production

Until J. P. Guilford introduced the operation of divergent production in

his Structure of Intellect (SOI), creativity was generally considered a phe-
nomenon separated from intelligence, a state of mind that was common for
those considered gifted and a blessing for those we perceive of as lucky.
Until Guilford, the prominent works could be roughly summarized to three:
Catherine Cox (Cox 1926), who argued that creativity was a complex, mul-
tivariate behavior (as opposed to a single ability or trait)3 ; Helmholtz (Von
Helmholtz 1896) and Wallas (Wallas 1926), the latter two being the creators
of the four steps model4 (preparation, incubation, illumination and verifica-
tion) that became the classical stance, within Psychology, of what a creativity
model should involve. This model is not contradicted by Guilford and still
concurs with many current views of the subject. In section 1.3, we will take
a closer look to this model. For the moment, we are interested in giving the
reader a short overview of SOI, with particular attention to divergent produc-
tion, the operation most linked with creative production.
The major aim of SOI was to give to “the concept of ‘intelligence’ a firm,
comprehensive, and systematic theoretical foundation” (Guilford 1967). This
very ambitious goal must be viewed from a historical perspective: during the
first part of the twentieth century, many measuring tests of mental ability ap-
peared, often motivated by the need to quantify “intelligence”. This need was
increased by the advents of the first and second world wars, when fast and
effective processes of selection were fundamental for recruitment (mainly in
areas such as the air force or intelligence services). The overstated relevance
of testing the concept of intelligence justified the sarcastic sentence of E. G.
Boring: “. . . intelligence as a measurable capacity must at the start be defined
as the capacity to do well in an intelligence test” (Boring 1923). Guilford
himself, who also made significant contributions to this area of psychome-
try, pointed out the lack of a coherent psychological theory behind tests in
 Creativity theories 9

general, this becoming the general motivation for the SOI. His more specific
intentions were to provide SOI as a frame of reference for the study of the
“intellectual factors”. An intellectual factor corresponds to an aspect of intel-
ligence represented by a triple operation/product/content (see figure 1).
Contents
Visual
Auditory
Symbolic
Semantic
Behavioral
Products
Units
Classes
Relations

Systems
Transformations
Implications

Operations
Evaluation
Convergent Production
Divergent Production
Memory
Cognition

Figure 1. The Structure of Intellect (SOI) (from (Guilford 1967))

For each factor, Guilford proposes tests, the majority of them implying
correlations of many factors. For example, for cognition of symbolic classes
(CSC), he suggests tests like:

Which pair of numbers does not belong with the others?

A. 1-5
B. 2-6
C. 5-8
D. 3-7
answer:C (in all other cases the difference is 4)
10 Creativity

This test has correlations with CSR (cognition of symbolic relations) and
CMS (cognition of semantic systems). With SOI, there would be no single
value to represent the intelligence of a subject, instead a set of values would
represent his/her main intellectual qualities and defects.
Perhaps the major contribution of SOI to the area of psychology and, more
specifically, to the notion of intelligence has to do with the demonstration,
supported by a large variety and quantity of empirical work, that intelligence
is not monolithic: there is a multitude of factors to take into account and one
cannot find a unique and absolute measure (as Catherine Cox had claimed
before). However uncontroversial and obvious this may seem today, the fact
is that only during the last decades of the twentieth century was there gen-
eral acceptance of the idea that single measures like the IQ are very fragile
indicators for peoples’s behavior and abilities. Another important contribu-
tion, the one that most interests us, has to do with the inclusion of creativity
as a fundamental aspect of intelligence. More specifically, Guilford considers
creative production as a general ability that humans have, and which depends
on many different intellectual factors, but, most of all, on an operation: that
of divergent production (DP). His formal definition of divergent production
reads: “generation of information from given information, where the empha-
sis is upon variety and quantity of output from the same source; likely to
involve transfer.” (Guilford 1967). DP composes four fundamental abilities:
– fluency - generation of a large number of solutions for a problem
– flexibility - generation of varied solutions
– originality - generation of solutions that are: rare within the population;
remotely related; clever responses
– elaboration - ability to think of details
As with the rest of SOI, Guilford proposes a series of tests. In DP-tests,
subjects are asked to exhibit evidence of divergent production in several ar-
eas, including that of semantic units (e.g. listing consequences of people no
longer needing to sleep), of figural classes (finding as many classifications
of sets of figures as is possible), and of figural units (taking a simple shape
such as a circle and elaborating upon it as much as possible). For example,
the following test should measure divergent production of semantic classes
(DMC):

From the list of words to the left, make some small subclasses of objects:
1. arrow alternate classes
2. bee 1,2,5,7 (found in the air)
 Creativity theories 11

3. crocodile 3,4,6 (found in the water)

4. fish 2,3,4,7 (animals)
5. kite 3,4,5,7 (have tails)
6. sailboat etc.
7. sparrow

From these tests and reflections on the whole model, Guilford also pro-
poses another concept as fundamental to creativity, that of transfer recall:
“Things are recalled in connection with cues with which they were not expe-
rienced before. Transfer recall is retrieval of information instigated by cues in
connection with which the information was not committed to memory stor-
age.” (Guilford 1967). In other words, transfer recall is the operation that
allows knowledge in memory, however semantically distant and apparently
unrelated to the problem at hand, to be brought and applied to a current situ-
ation. This is what we call cross-domain transfer throughout this book.
To summarize, the operation of divergent production is the very basis for
the set of phenomena that are commonly associated with creativity in people,
although, as Guilford himself points out,

‘...creative potential is not a single variable, any more than

intelligence. Creative performances in daily life are enormously
varied in the demands that they make on intellectual resources.
The performances singled out for their more obvious signs of
creativity - novelty, ingenuity, inventiveness - probably involve
one or more divergent production abilities as key aspects, or
transformation abilities, outside the DP-operation category as
well as within it.’

Although creativity is normally linked more to free-association, uncon-

strained reasoning or unexpectedness than to method, constraint satisfaction
or inference, it has been clear from reading many studies (many described
or referred to within this document) that a great deal of mastery of knowl-
edge, expertise within a domain and focus is fundamental. Thus, although
not so much emphasized by Guilford, the converse operation of DP, conver-
gent production (CP), is also fundamental (as pointed out by Csikszentmi-
halyi, in section 1.3), since it provides deductive reasoning or compelling
inferences. “Convergent production rather than divergent production is the
prevailing function when the input information is sufficient to determine a
unique answer.” SOI tests for evaluating convergent production essentially
12 Creativity

measure the ability to solve puzzles, equations, classification tasks and prob-
lems in general that yield a logically sound unique solution.
Guilford makes a comparison between DP and CP:
‘[In DP], the problem itself may be loose and broad in the re-
quirements for solutions; or the problem, if properly structured,
may call for a unique solution, but the individual may have an
incomplete grasp of it; or he may comprehend the problem fully,
but he is unable to find the unique answer immediately.(..) In CP,
an answer can be rigorously structured and is so structured and
an answer is forthcoming without much hesitation. In the former,
restrictions are few; in the latter they are many; in the former, the
search is broad; in the latter it is narrow. In the former, criteria
for success are vague and somewhat lax and may, indeed, stress
variety and quantity; in the latter, criteria are sharper, more rig-
orous, and demanding.’
Thus, according to Guilford, CP and DP are two complementary facets of
our productive capacity. This capacity, along with cognition (which he con-
siders a more specific operation: that of comprehension and understanding),
memory and evaluation, make part of a model of problem solving and creative
production that the author proposes as an operational integration of all the as-
pects of SOI. Although this model is essentially a speculation, it is interesting
to reproduce the original diagram to the reader (figure 2).
It is far outside the scope of this book to present this model in detail. Since
Guilford did not explore it deeply himself, it is an abstract suggestion for how
things should be when solving problems. Two aspects, however, should be re-
tained: Guilford argued for problem solving and creative production being the
same, thus building a common model for both; he considers a heterogenous
memory with many kinds of representation, perspectives, domains, all co-
habiting together, in an organized whole. This is what we call multi-domain
environment throughout this book.
We would like to finish this section with some thoughts about Guilford’s
work, taking into account, obviously, that this is a work that is almost 40
years old. The first issue is the supremacy of verbal versus non-verbal repre-
sentation, which sometimes seems to imply that thought is defined by (verbal)
language and not the opposite. Allowing some speculation from our side, we
believe this is due to the dualist tendency of the time, where mind is detached
from matter, as opposed to the current view, of embodiment, where some
 Creativity theories 13

Figure 2. Guilford’s model of problem solving and creative production (from (Guil-
ford 1967: 315-316)

cognitive scientists see consciousness and mind as a result of the interaction

of cognition with the whole physical experience. In other words, today, Guil-
ford’s symbolic and semantic contents would be connected as much with ver-
bal symbols as with any other kinds of symbols, such as social or behavioral
symbol systems.
Another criticism relates to the vagueness of some definitions, such as
cleverness or generation of logical possibilities. These being related to diver-
gent production, it is very important to clarify them. Here too, we must al-
low ourselves some speculation bounded by context. By cleverness, or clever
solutions to a problem, the author means solutions that both respond more
effectively to a problem (than the usual, convergent, ones) and are rare to find
in a search space. When defining divergent production (as opposed to con-
vergent) as being with the ability to generate logical possibilities, Guilford
meant the creation of unsound logical facts (or rules) that, however, are not
revealed to be inconsistent with the rest of the knowledge (e.g. facts that are
not deducible, but do not contradict existing knowledge). Here, too, we would
like to add that sometimes inconsistencies do arise and creativity comes out
14 Creativity

of the confrontation between the inconsistencies and the theory (e.g. Kepler’s
discoveries about elliptic versus circular orbits5 ). Guilford would certainly
agree with this opinion.
In conclusion, the Structure of Intellect is now certainly outdated, and its
contribution is now seen from a historical perspective. However, with regard
to the psychology of creativity, Guilford’s legacy about divergent and conver-
gent production is still a constant reference. For us, it has become a modelling
inspiration.

1.2. Bisociation

Before going into further detail about Arthur Koestler’s work, we would like
to add that The Act of Creation is a rich philosophical and psychological ref-
erence that encompasses ideas that are still currently accepted and explored,
as we can see in current trends such as embodiment or conceptual blending.
In many ways, bisociation prefigures the blending framework. Indeed, we
might even ask in what way is blending merely “parameterized bisociation”,
that is, bisociation with more elaborate principles and constraints. This will
be discussed in due course.
Written during the early sixties, when behaviorism was the dominating
trend within psychology, this book takes the opposite position (which is close
to a structuralist view) and aims to explain the act of creativity, tackling this
challenge from several different perspectives. While it certainly misses many
aspects and perhaps fails in depth to favor the breadth, it proposes a set of
ideas that we will try to synthesize here and which will be taken by us for the
sake of argument of some of our options.
In The Act of Creation, Arthur Koestler presents a theory that unifies three
sides of human behavior commonly deemed creative: humor, science and the
arts. According to him, the underlying processes are the same, but applied
in different contexts, subject to different intentions and perspectives. In order
to support his theory, Koestler proposes a set of definitions regarding knowl-
edge and problem solving, namely matrices of thought, codes of rules and
strategies.
A matrix of thought (or simply, a matrix) is “any ability, habit, or skill,
any pattern of ordered behavior governed by a ‘code’ of fixed rules” (Koestler
1964). In his example of chess playing (as an ability), he proposes a matrix
as being “the pattern before you, representing the ensemble of permissible
 Creativity theories 15

moves. The code which governs the matrix can be put into simple mechanical
equations which contain the essence of the pattern in a compressed, ‘coded’
form”. A code of rules is then what defines the matrix, which means that
both represent the same entity from different perspectives, one intensional,
the other extensional. A strategy corresponds to the selection of elements
within the matrix in order to achieve a goal or pattern of behavior. In the case
of chess, this would be the choice of the “next move”.
We find several obvious correlations with AI problem solving, namely a
matrix corresponds to the set of all possible solutions for a given problem (the
solution space), defined by the code of rules, a set of constraints that define
what a valid solution must be like. The strategy is then the search procedure,
the method used to choose solutions.
According to Koestler, the creative process is connected to what he terms
bisociation of matrices, a phenomenon that occurs when two (or more) ma-
trices become intersected: a reasoning is being followed in one of the matri-
ces and, for some reason (e.g. external stimulus, need, dream, or trance-like
state), a “clash” happens with another matrix and there is a leap to an alternate
reality.

1.2.1. Humor

In humor, for example, bisociations introduce the sudden association to

the unexpected, the illogical that triggers laughter, sometimes via double-
meaning, phonetics, caricature, satire, to name but a few. Let us quote a short
example:
Two women meet while shopping at the supermarket in the
Bronx.
One looks cheerful, the other depressed. The cheerful one
inquires:
’What’s eating you?’
’Nothing’s eating me’
’Death in the family?’
’No, God forbid!’
’Worried about money?’
’No...nothing like that.’
’Trouble with the kids?’
’Well, if you must know, it’s my little Jimmy.’
16 Creativity

’What’s wrong with him, then?’

’Nothing is wrong. His teacher said he must see a psychiatrist.’
Pause. ’Well, well, what’s wrong with seeing a psychiatrist?’
’Nothing is wrong. The psychiatrist said he’s got an Oedipus
complex.’
Pause. ’Well, well, Oedipus or Shmoedipus, I wouldn’t worry so
long as he’s a good boy and loves his mamma.’

Here, we see a clash between the matrices of Freudian psychiatry patholo-

gies and the logic of common sense: if Jimmy is a good boy and loves his
mamma, there can’t be much wrong. Koestler even arrives to an extreme
claim that “any two matrices can be made to yield a comic effect of sorts, by
finding an appropriate link between them and infusing a drop of adrenalin”
(Koestler 1964).

1.2.2. Science

In his extensive analysis of scientific discovery throughout the history of

science, Koestler points out several observations that converge to the idea
that “each basic advance was effected by a more or less abrupt and dramatic
change; the breaking down of frontiers between related territories, the amal-
gamation of previously separated frames of reference or experimental tech-
niques (..) All decisive advances in the history of scientific thought can be
described in terms of mental cross-fertilization between different disciplines”
(Koestler 1964). In other words, all the major advances observed resulted
from bisociative thinking and these include examples like the discoveries
from Archimedes, Copernicus, Kepler, Galileo, Darwin, Poincaré, Kekulé,
Einstein, to name a few. In summary, in each of these situations, the knowl-
edge has so far proved inapplicable; none of the various ways of exercising a
skill, however flexible and adaptable, has led to the desired goal. The solution
came out of a new synthesis of previously unconnected matrices of thought; a
synthesis arrived at by “thinking aside”. For example, Kepler’s laws of plan-
etary motion represent the first synthesis of astronomy and physics which,
during the preceding two thousand years, had developed along separate lines.
Kepler served his apprenticeship under Tycho de Brahe, who had improved
the astronomy observation instruments and methodology, thereby allowing
hitherto unequalled abundance and precision. Given the new data, there were
 Creativity theories 17

clear inconsistencies in the traditional astronomy predictions, mainly because

they were based on very entrenched dogmas such as that “all heavenly mo-
tion must be uniform and in perfect circles”. By retaining much of the meta-
physical and theological basis that Kepler himself believed in, he was able
to postulate the existence of a physical force acting between the sun and the
planets, thus leading to a revolution in astronomy. Planets no longer move in
circles, but in elliptic orbits.
Koestler also points out the fundamental aspects of the ripeness of the dis-
coverer, i.e. he must be prepared, predisposed to the discovery6 , and have the
ability to find hidden analogies (i.e. to find relations where no one has found
them before) within different matrices. Another interesting observation is that
“verbal thinking plays only a subordinate part in the decisive phase of the
creative act (..) as the creative process of discovery depends on unconscious
resources and presupposes a regression to modes of ideation which are indif-
ferent to the rules of verbal logic”. The conclusion is that words are essential,
but sometimes become snares, decoys, or strait-jackets.
An interesting quotation comes from the mathematician Henri Poincaré:

Among chosen combinations the most fertile will often be

those formed of elements drawn from domains which are far
apart... Most combinations so formed would be entirely sterile;
but certain among them, very rare, are the most fruitful of all.

1.2.3. Arts

The third perspective that Koestler analyzes is that of the artist. As with hu-
mor, he starts by focussing on the physical manifestations connected to it,
the many sensuous phenomena we feel in moments of self-transcendence,
from goose bumps to weeping, thereby arriving at the emotive potential of a
matrix, with its capacity to generate and satisfy participatory emotions (e.g.
by identification, aggression, tension, relaxation). Perceiving a form of art,
of deluding oneself without losing track of reality, means exploring this ma-
trix with higher emotive potential, moving from the trivial present to a plane
remote from self-interest while forgetting current preoccupations and anxi-
eties: “The capacity to regress, more or less at will, to the games of the un-
derground, without losing contact with surface, seems to be the essence of the
poetic, and of any other form of creativity”. The author thus proposes the act
18 Creativity

of interpretation as being also bisociative, and thus creative from the point of
view of the recipient.
As in scientific discovery, metaphor and imagery also come into existence
by a process of seeing an analogy where no one saw one before, its aesthetic
satisfaction depending on the emotive potential of the matrices involved. Ac-
cording to Koestler, discoveries of art derive from “the sudden transfer of
attention from one matrix to another with a higher emotive potential”. In
other words, as with science, the greatness of an artist rests in creating a new
idiom - a novel code which deviates from the conventional rules. The key
turning points result from a new departure along a new line, where we can
find bisociations in the grand style - cross-fertilization between different pe-
riods, cultures, and provinces of knowledge. Once a new idiom is established,
“a whole host of pupils and imitators can operate it with varying degrees of
strategic skill”. Here, Koestler clearly shows his view on “true creativity -
the invention of a new recipe” as opposed to “the skilled routine of providing
variations for it”. This dichotomy also corresponds to the transformational
and exploratory creativity that Margaret Boden discusses, which will be ap-
proached in section 1.4.
In conclusion, Koestler argues that bisociation is active in those three as-
pects of human creativity. In humour, by the collision of matrices; in science
by their integration; and in arts by their juxtaposition.
Although rarely showing it in a formal or algorithmic fashion, Koestler
provides an insight to one of the most definite phenomena behind the creative
process, that of combining knowledge from different domains. More than
describing in detail what happens in cognition, he identifies the consistent
existence of what he calls bisociation within a very wide range of situations
commonly deemed creative, and while it may be arguable that not every cre-
ative act complies with this description, it is certainly true that many of them
result from the association of apparently semantically distant sources, which,
for some reason, combine in such a way that novel and useful knowledge
emerge so naturally that those sources no longer seem so distant. From the
perspective of Guilford’s theory, the notion of bisociation takes many inter-
sections with that of transfer recall, although the latter seems reduced to the
act of retrieving unexpected elements from memory, and the former to the
actual processes of combination involved.
 Creativity theories 19

1.3. The Systems Model

The work of Mihaly Csikszentmihalyi on Creativity (Csikszentmihalyi 1996)

is the most recent of the ones reviewed in detail here. This does not mean,
however, that it has a radically different perspective as we will see. Indeed
one of the purposes of looking at this widely referred study is to demonstrate
the current validity and acceptance of the works of Koestler and Guilford, as
well as to provide some other comments that we believe are of importance to
the present work.
From a study that took several years, in which the author and colleagues
interviewed and analyzed ninety one people widely deemed creative, Csik-
szentmihalyi proposes a description of how creativity works and how culture
evolves as domains are transformed by the curiosity and dedication of a few
individuals. He proposes a systems model of creativity that takes into account
the interaction of three elements: the domain, the person and the field. The
domain consists of a set of “symbolic rules and procedures. Mathematics is a
domain, or at a finer resolution algebra and number theory can be seen as do-
mains. Domains are in turn nested in what we usually call culture, or the sym-
bolic knowledge shared by a particular society, or by humanity as a whole.”
(Csikszentmihalyi 1996) Csikszentmihalyi defines the creative person as des-
ignating individuals who, like Leonardo, Edison, Einstein or Mozart have
changed our culture in some important respect. The field includes all the in-
dividuals and institutions that act as gatekeepers to the domain, the peers that
will judge the person and the ideas. It is the field that selects what new works
of art, objects or theories deserve to be recognized, preserved and remem-
bered. Each of these three elements is necessary for a creative idea, product,
or discovery to take place.
Thus the author proposes a very strong definition of creativity. Let us call
it creativity with a big C: “Creativity is any act, idea, or product that changes
an existing domain, or that transforms an existing domain into a new one (..)
A creative person is someone whose thoughts or actions change a domain, or
establish a new domain (..) So, in a sense, the most momentous creative events
are those in which entire new symbolic systems are created.” (Csikszentmi-
halyi 1996). In order to understand the process behind these transformations
that shape civilization, Csikszentmihalyi analyzed and interviewed individu-
als who were able to create or drastically change one domain, at least once in
their lives. The list included renowned artists, politicians, economists and sci-
entists, some having been awarded Nobel prizes or other important awards in
20 Creativity

their field. Although presumably not providing statistical validity (e.g. there
is no control group or objective measures in general), the author identified
a set of traits that were common among the individuals (Csikszentmihalyi
1996):

1. Creative individuals have a great deal of physical energy, but they also
know how to be quiet and at rest.

2. Creative individuals tend to be smart, yet also naive at the same time.

3. A third paradoxical trait refers to the related combination of playful-

ness and discipline, or responsibility and irresponsibility

4. Creative individuals alternate between imagination and fantasy at one

end, and a rooted sense of reality at the other. Both are needed to break
away from the present without losing touch with the past.

5. Creative people seem to harbor opposite tendencies on the continuum

between extroversion and introversion.

6. Creative individuals are also remarkably humble and proud at the same
time.

7. In all cultures, men are brought up to be masculine and to disregard

and repress those aspects of their temperament that the culture regards
as feminine, whereas women are expected to do the opposite. Creative
individuals to a certain extent escape this rigid gender role stereotyping.

8. Generally, creative people are thought to be rebellious and independent.

Yet it is impossible to be creative without having first internalized a
domain of culture (...) hence it is difficult to see how a person can be
creative without being both traditional and conservative and at the same
time rebellious and iconoclastic.

9. Most creative persons are very passionate about their work, yet they
can be extremely objective about it as well.

10. Finally, the openness and sensitivity of creative individuals often ex-
poses them to suffering and pain yet also a to great deal of enjoyment.

On his explorations concerning the creative process itself, Csikszentmi-

halyi follows the traditional description of five steps (that descends from the
 Creativity theories 21

four step model of Wallas, with Verification split into Evaluation and Elabo-
ration): Preparation, Incubation, Insight, Evaluation and Elaboration. He then
frames these steps within the systems model of domain, person and field.
As confirmed in each of the individuals studied, there is the need for a
tremendous amount of information about the domain and the field, normally
achieved after years of hard work. This preparation depends as much on ex-
ternal factors (e.g. family, education, socioeconomic factors, political issues
within a field) as on internal factors (e.g. curiosity, persistence, talent). The
incubation is generally described as the process of solving a problem in the
underground of cognition, after the creative person feels that he/she have been
blocked: “Cognitive theorists believe that ideas, when deprived of conscious
direction, follow simple laws of association. They combine more or less ran-
domly, although seemingly irrelevant associations between ideas may occur
as a result of a prior connection”(Csikszentmihalyi 1996). This mysterious
and often controversial step was also confirmed by the individuals studied,
who often reported finding the solution to a problem in unpredictable or con-
sciously unprepared situations, sometimes after years of working on the prob-
lem (or even having left it alone). As the author points out, this insight, also
found recurrently in the literature of creativity, is only possible when the per-
son (in fact the whole system) is prepared to identify it. This corresponds to
the idea of ripeness, as described by Koestler. For this to happen, the per-
son must be in the right place at the right time, with a significant amount of
confidence, knowledge and luck, as frequently confirmed by the interviewed
individuals.
Evaluation and elaboration are steps that gradually become more depen-
dent on the whole system and less on the individual, as the act, idea, or prod-
uct is confronted with the domain and the field, although the person also
becomes part of this evaluation, particularly in less objective domains.
Two important aspects which raise from Csikszentmihalyi’s observations
are the duality of divergent/convergent thinking and integration across and
within domains, both of which are consistently reported and analyzed. From
the observations, creative people are able to perform well (and with constant
switches) in both the opposite ways of thinking reported in section 1.1: diver-
gent and convergent.

‘People who bring about an acceptable novelty in a domain seem able to use
well two opposite ways of thinking: the convergent and divergent(..). Diver-
gent thinking is not much use without the ability to tell a good idea from a
22 Creativity

bad one - and this selectivity involves convergent thinking.’

Divergent thinking is extremely important for the phases of Preparation

and Incubation, since these phases are characterized by curiosity and search-
ing. Convergent thinking is determinant for the Elaboration, Evaluation and
also Insight, since it brings about the ability to tell a good idea from a bad
one, to follow the established rules of the domain and confront them with new
knowledge. However, the boundaries must not be so strict. Indeed, the cre-
ative process intertwines frequently between divergent and convergent think-
ing, as well as any of the five steps just described.
In the same way that Guilford’s contribution (divergent and convergent
thinking) has been accepted and backed up by the work of Csikszentmihalyi,
so Koestler’s bisociation also appears as “the norm rather than the exception”.
Although rarely identifying the phenomenon with the name of bisociation,
the author repeatedly provides examples and reports of situations that involve
the cross-domain transfer of ideas, the bringing together of domains that ap-
pear to have nothing in common and integration or synthesis both across and
within domains. As he points out, “creativity generally involves crossing the
boundaries of domains, so that, for instance, a chemist who adopts quantum
mechanics from physics and applies it to molecular bonds can make a more
substantive contribution to chemistry than one who stays exclusively within
the bounds of chemistry.” (Csikszentmihalyi 1996).
The work just covered suggests many important issues for the modelling
of creativity and of creativity supporting tools. From a system’s perspective,
one should not simply focus on one single element (person, domain or field),
for it is from their interaction that creativity emerges. It also raises many traits
and conditions that need to be present for the creation of novel and useful
ideas. However we must point out that this study lacks many of the scientific
bases necessary to assert with much confidence some of these more abstract
conclusions, namely because the individuals come from a very specific class:
successful, recognized, well established people, in general happy with their
own accomplishments, and normally in an advanced phase of their lives (over
sixty years old, in general). This means that many creative and non-creative
people have been left out, some of whom are unsuccessful so far, struggling
to be recognized, yet still be extremely creative. In other words, the set of in-
dividuals corresponds to the class of creative persons that the system (person,
domain and field) has brought to the world, but the system is too dynamic not
to confront data with other systems (e.g. a control group) or other states of
 Creativity theories 23

the system (e.g. different ages).

Another aspect that Csikszentmihalyi himself has raised is that, according
to these criteria, children can be talented but never really creative, because
“creativity involves changing a way of doing things, or a way of thinking,
and that in turn requires having mastered the old ways of doing or think-
ing.” He thus leaves space for two other classifications: personal creativity,
experiencing the world in novel and original ways; and talent or brilliance,
the ability to express unusual thoughts which are interesting and stimulating.
Any of these three classifications associated with the word “creativity” (true
creativity, personal creativity and talent) provides interesting challenges to
computational modelling and also leads us to Boden’s taxonomy (h-creativity
and p-creativity), thus providing similar analyses with regard to AI and com-
putational creativity.

1.4. Boden’s taxonomies

Of the works and authors in this section, there is no doubt that Margaret
Boden (Boden 1990) is the most read and cited within the field of AI and
Creativity. The simple reason for this is the fact that she pioneered the effort
of analyzing some of the work that had been done (in AI) from a perspective
that takes into account those philosophical and psychological issues which
are traditionally deemed as creative. In so doing, she proposes a set of classi-
fications for analyzing a program (as much as a human). These classifications
themselves have raised much debate, some of which we will cover here.
The first classification proposed by Boden concerns the fact that there is
novelty in an idea or discovery: Whether it is novel for a single person or for
the whole of human history. In principle, every new and useful idea or dis-
covery is creative for its producer. This is called psychological creativity (or
p-creativity). The other reference is history. When an idea is novel and useful
for the whole of human history, then we are faced with historical creativity
(or h-creativity). This is perhaps the less controversial classification, although
many authors argue that true creativity cannot exist without an external eval-
uation (e.g. (Csikszentmihalyi 1996), (Lubart 1999)). Another aspect is that
one can never determine h-creativity in absolute terms because an idea can
be h-creative and not be seen as such for years. And the reverse also hap-
pen, when an idea and an author are regarded as h-creative by the society,
24 Creativity

but the original idea should actually have been credited to a preceding author.
There are countless examples of these misjudgments in human history. This
view does not remove the validity of h-creativity as presented, but it testifies
to how complex the problem can be. It is also important to mention that p-
creativity is the main focus of Boden’s analysis, as she is mainly concerned
with the personal perspective.
The other classification brought by Boden (Boden 1990) pertains to the
process needed to produce the novel idea or discovery. She thus presents two
kinds: combinatorial creativity and exploratory-transformational creativity.
The combinatorial creativity results from “unusual combination of, or associ-
ation between, familiar ideas. Poetic imagery, metaphor and analogy fall into
this class.” (Boden 1999). Alternatively, exploratory-transformational cre-
ativity (ET-creativity) is about how a subject deals with a conceptual space.
Her definition of conceptual space is: “an accepted style of thinking in a par-
ticular domain - for instance, in mathematics or biology, in various kinds of
literature, or in the visual or performing arts...” (Boden 1999). There is a clear
similarity with Koestler’s matrices throughout the many descriptions that Bo-
den provides for conceptual space, although neither provide a formal defini-
tion. ET-creativity further subdivides into two distinct categories: exploratory
(e-) and transformational (t-). E-creativity deals with the exploration of the
conceptual space without jumping out of its boundaries, without breaking
strong constraints, and it is normally based on the mere “tweaking” of the
most superficial dimensions. Sometimes it is capable of achieving p-creative
or even h-creative outcomes, but still without changing the (well defined) con-
ceptual space. T-creativity involves some transformation “of one or more of
the (relatively fundamental) dimensions defining the conceptual space con-
cerned” (Boden 1999). In other words, it demands changes in the conceptual
space, such as re-representation, change in the evaluation or integration of
new concepts. Boden also sees this kind of creativity as impossibilist, “in that
ideas may be generated which - with respect to the particular conceptual space
concerned - could not have been generated before (they are made possible by
some transformation of the space).” (Boden 1990:519-520). Thus, there is a
clear opposition between e- and t- creativity, although for some authors, it is
about the level of abstraction. Indeed this taxonomy has raised many points
of debate.
The first point is that transformational creativity is also exploratory at a
meta-level (Wiggins 2001; Colton 2001; Ram et al. 1995). In other words,
given the nature of t-creativity, the only possible way to transform a concep-
 Creativity theories 25

tual space is to change its own defining rules. This would involve being aware
of its own defining rules, in other words, being able to do meta-level reason-
ing. Following this argument, it leads us to the conclusion that this change of
meta-level rules would necessarily be (at some point, even if at a meta-meta-
level, and so on) exploratory. This argument has been formalized in (Wiggins
2001).
Another criticism concerns the vagueness of the definition of conceptual
space (Wiggins 2001; Ritchie 2001). Although she provides many exam-
ples, such as from within the broad areas of expertise like music, writing
or physics, it is never sufficiently clear from a computational modelling per-
spective what it actually comprises. More specifically, should it correspond
to a solution set, i.e. the set of solutions to a problem? Is there any ordering,
so it becomes then a search space? This issue may become important when
considering the computational modelling (and analysis) of t- and e-creativity.
For example, changing the ordering of concepts would correspond to a trans-
formation of the search space, but not of the solution set. Since in the latter
case there is no introduction of new concepts, one cannot say it could not have
been generated before (therefore it should be e-creativity). On the other hand,
many discoveries and art revolutions (i.e. h-creative events) may have been
based more on this kind of restructuring the space than on the generation of
impossible concepts, which would mean that t-creativity is not necessarily a
superior kind of creativity. We think that these issues could be better clarified
with a more precise definition of conceptual space.
The final point to present here regards the distinction between combinato-
rial creativity and ET-creativity (Ritchie 2001). Although not having raised
as much debate as the previous issues, essentially because Boden herself
dropped this differentiation, this one has particular interest for our work. If
one sees the problem of combinatorial creativity as the generation of a new
concept from the association of previous ones (as the definition says), one
can also accept a conceptual space containing all the possible combinations.
By doing so, there is no difference between the act of exploring the concep-
tual space of possible combinations, and the act of generating a combination
(which would exist in that conceptual space). Similarly, if, with the novel as-
sociation, a novel concept emerges that could not have been generated before,
then we could have achieved t-creativity. In other words, although combina-
torial creativity may be regarded as a particular kind of creativity (which is
also the one approached in this book), it should also be included in and not
distinguished from the set of ET-creativity phenomena. This is, as far as we
26 Creativity

know, a common and uncontroversial perspective towards ET- and combina-

torial creativity.
It is without doubt that Margaret Boden produced a comprehensive analy-
sis of creativity and AI that has been used and applied in many works. We too
will use some of the ideas described. Above all, she contributed successfully
to the provocative question about whether computers can be creative. Indeed
they can be, although perhaps at a very limited level in comparison to human
creativity, at the least at a level that does not demand self-awareness. This
may be a very mechanistic, unromantic, level, but it is nonetheless clearly
able to surprise humans with outcomes that we ourselves would normally
have no problem in considering as the result of creative behavior. In section
2, we will give an overview of some of these systems.

1.5. Others

Given the recent increase in creativity research, it is not a simple task to pro-
vide a meticulous overview without leaving out any fundamental work. There
is a great variety of approaches and therefore we will provide a summary of
the most cited works from those approaches that are most prominent and
field-covering: cognitive psychology (Finke, Ward and Smith 1992); conflu-
ence theories7 (Sternberg and Lubart 1996); neuroscience (Martindale and
Greenough 1974); motivation and intention (Amabile 1983) and biographical
case studies (Weisberg 1999).
Finke and his colleagues have proposed what they call the Geneplore
model, according to which there are two main phases in creative thought:
generative and exploratory (Finke, Ward and Smith 1992). Many potential
ideas or solutions are created, followed by their extensive exploration. From
laboratory experiments, these researchers concluded that subjects generate a
set of “preinventive structures, in the sense that they are not complete plans
for some new product, tested solutions to vexing problems, or accurate an-
swers to difficult puzzles” (Ward, Smith and Finke 1999). From these partial
structures, a phase of exploration and interpretation takes place that attempts
to construct a feasible solution to a problem, by focusing and expanding these
structures. Constraints can be imposed at any time during the generative or
exploratory phase. This model “acknowledges that a range of factors other
than cognitive processes contribute to the likelihood of any individual gener-
ating a tangible product that would be judged to be ‘creative’ ” (Ward, Smith
 Creativity theories 27

and Finke 1999). From a broad perspective, the Geneplore model falls into
the class of divergent-convergent models, as proposed by Guilford and agreed
by Csikszentmihalyi.
The investment theory of Sternberg and Lubart falls into the category of
confluence theories (theories that offer the possibility of accounting for di-
verse aspects of creativity). It suggests that “creative people are ones who are
willing and able to ‘buy low and sell high’ in the realm of ideas. Buying low
means pursuing ideas that are unknown or out of favor but that have growth
potential. (..) The person persists in the face of this resistance and eventually
sells high” (Sternberg and Lubart 1996). From extensive experimentation,
Sternberg and Lubart developed a model that presupposes the interaction of
six distinct but interrelated resources: intellectual abilities, knowledge, styles
of thinking, personality, motivation, and environment.
At a different level of research, Martindale and Greenough studied the
variability of level of arousal and attention in the performance of creativity
tests (e.g. Remote Associations Test, Similarities Test of divergent thinking),
by observing galvanic skin response fluctuations, heart rate variability, corti-
cal activation, as well as other biometrical measures (Martindale and Gree-
nough 1974; Martindale 1999). One interesting conclusion was that creative
individuals have a highly variable level of arousal, rather than a basal (i.e.
stable in this context) level of arousal, which means that “creative inspiration
occurs in a mental state where attention is defocused, thought is associative,
and a large number of mental representations are simultaneously activated”
(Martindale 1999). Moreover, the authors also associate their work with the
primary-secondary process thesis (Kris 1952), which says that creative indi-
viduals have a greater ability to switch between two modes of thought (pri-
mary and secondary) than less creative individuals. Primary process thought
is found in states such as dreaming and reverie (as well as in psychosis and
hypnosis), it is autistic, free-associative, analogical. Secondary process is the
abstract, logical, reality-oriented thought. Martindale found supportive evi-
dence confirming Kris’s proposal, meaning that creativity is not just based
on primary process thought, but on its systematic intertwining with the sec-
ondary process. Again, we find a clear similarity to the ideas of divergent
thinking (the primary process) and convergent thinking (the secondary pro-
cess), as discussed before.
The work of Teresa Amabile on motivation and intention is also often cited
in literature. She proposes a two-pronged hypothesis about how motivation
affects creativity: “The intrinsically motivated state is conducive to creativity,
28 Creativity

whereas the extrinsically motivated state is detrimental” (Amabile 1983). In-

trinsic motivation is associated with the enjoyment of the work in itself, while
extrinsic relates to engaging in an activity in order to meet some goals exter-
nal to the work. She also proposes a confluence model with intrinsic motiva-
tions, domain-relevant knowledge and abilities, and creativity-relevant skills.
These include: “ a) cognitive style that involves coping with complexities and
breaking one’s mental set during problem solving; b) knowledge of heuristics
for generating novel ideas, such as trying a counterintuitive approach, and c)
a work style characterized by concentrated effort, an ability to set aside prob-
lems, and high energy (..)”.
The issue of re-representation is emphasized by some researchers (e.g.
(Karmiloff-Smith 1993; Oxman 1997)), who propose that a process of rep-
resentational re-description precedes creative domain exploration, in which
previously implicit knowledge is more clearly perceived. New patterns thus
emerge and novel inter-domain connections may be made.
Finally, in a totally different direction (essentially based on biographical
case studies), Robert Weisberg challenges many of the works just described
above by arguing against what he calls the tension view of the relationship
between creativity and knowledge (Weisberg 1999), which says that: since
knowledge about a problem is not complete, and in face of a blockage, the
person is left to its abilities to discover new solutions via free-association,
divergent thinking, etc. This discontinuous view of knowledge evolution is
a “dominant one in modern theory” (Weisberg 1999). Weisberg proposes a
continuous view, by arguing that new discoveries and revolutionary artworks
are the result of a state of maturity and knowledge richness. This foundation
view thus concludes that creativity and knowledge are positively related: “The
reason that one person produced some innovation, while another person did
not, may be due to nothing more than the fact that the former knew something
that the latter did not.” (Weisberg 1999).
As pointed out at the beginning, there are certainly important works that
have been left out. Yet, for the present book it is more important to provide the
reader with its fundamental bases and their synthesis than with giving exhaus-
tive knowledge about lateral issues. The idea here was two-fold: to suggest
that, from a computational modelling perspective, many different approaches
can be taken; but, nonetheless, there are common aspects across almost all of
the approaches. We will attempt to include these aspects in the next section.
 Creativity theories 29

1.6. Synthesis

Amongst the works mentioned here, there is an undisputed agreement that

creativity involves the creation of a novel and useful product. We may find dif-
ferent words for creation (generation, production), novelty (originality), use-
fulness (value, appropriateness, utility, significance, adaptability) and product
(idea, concept, solution to a problem), but there is no doubt that these are ei-
ther synonymous or different perspectives on the same subject. Even though
not as uncontroversial, the majority of the works also agree that the cognitive
processes that bring about a novel and useful product consist of a pair of op-
posite styles of thinking: divergent and convergent. The former is character-
ized by allowing ideas that defy logical reasoning (e.g. unsound conclusions,
contradictory associations, inconsistent sets of facts) or that correspond to un-
precedented associations8 . In opposition, convergent thinking corresponds to
logical reasoning, which follows well-defined constraints and is normally as-
sociated to methodic, purposeful thought9 . We have already presented these
two concepts in section 1.1, but, as often equalled, they have been associated,
when not synonymous, to primary/secondary-process (Kris 1952) and gener-
ative/exploratory (Finke, Ward and Smith 1992) (also, to a lesser degree, to
bisociation (Koestler 1964), transformational/exploratory (Boden 1990) and
incubation/elaboration (Wallas 1926), and therefore needed to be put in a
more actual context. The other motivation for this redefinition is to claim
that, if a creative product must be novel and useful and its creation involves
the process of divergence and convergence, then one cannot build a creative
system without modelling both divergent and convergent processes. More-
over, divergence would be primarily responsible for granting novelty, and
convergence for usefulness. We will come back to these issues later.
It has also been repeatedly emphasized that knowledge is of central im-
portance in creativity. Virtually every author in this study argued that no im-
portant discovery or major artwork is likely to transpire without its creator
having acquired deep knowledge about the domain in question. Furthermore,
some also claim that having broad knowledge is also fundamental, in order
to potentiate transfer across domains and awareness of the environment. Here
we have described two perspectives: in-depth and in-breadth. Again, we are
tempted to associate these two categories with convergence and divergence,
respectively. This means that knowledge from the domain in question would
in turn benefit the convergence towards useful solutions, helping to discern
the good from the bad, while knowledge from a variety of domains would
30 Creativity

promote divergence in problem solving.

The existence of different levels of creativity has also been claimed by
many authors. Some arguing for a continuum (e.g. (Koestler 1964), some for a
clear distinction (e.g. (Boden 1990)): On one extreme, we have the creativity
with a big C, transformational creativity or true creativity. On the other side,
we have the personal, exploratory, or mundane creativity. We do not intend
to say that these are synonymous concepts, but to emphasize the bipolarity of
the analyses made.
A final and more controversial issue pertains to the role of society in cre-
ativity. Some authors (e.g. (Csikszentmihalyi 1996)) argue that there can be
no true creativity10 without the society, i.e., something does not exist as cre-
ative unless it is externally judged as so. It is a dynamic ascription that de-
pends on the interaction of several entities (in the case of Csikszentmihalyi,
the person, the domain and the field). Others (e.g. (Finke, Ward and Smith
1992)) argue for the existence of creativity for its own sake, i.e. an indi-
vidual can generate a creative idea, without having feedback from the so-
ciety. These two points of view correspond to what Boden called h- and p-
creativity, respectively, and thus reflect more a difference of perspective than
of the essence of creativity.

2. Computational Creativity

In this section, we introduce the area of computational creativity, which aims

to build computational systems that are able to exhibit creative behavior.
Since this is not a universally accepted definition (how can we know, after
all, when the behavior is creative?), some works carry on to the field of Arti-
ficial Intelligence the debate that we have presented in the previous section.
One of the immediate effects of this is the need for formal accounts for cre-
ativity, while the other is the experimental implementation of such systems.

2.1. Two accounts for characterizing Creativity

The legacy left by Boden’s descriptive hierarchy sparked off the attention to-
wards analyzing AI systems from a creativity perspective (e.g. (Ram et al.
1995; Bentley 1999)). Notwithstanding the many fragilities, some of which
have already been named, the point was made of the need of such analyses
 Computational Creativity 31

or, at the least, for the consideration of Creativity within AI. However, the
lack of formal approaches and of stable epistemological and methodological
settlements condemns this research to endlessly recycle unsolvable problems.
While it seems currently impossible to say that a system is creative, or even
intelligent, without any controversy, it may be possible to classify it accord-
ing to criteria that are based both on formal computational accounts and on
theories such as Boden’s.
We now present the two approaches towards characterizing the creativity
of AI systems. The first one derives from an assumed attempt of formal-
izing Boden’s theory and is centered on the process, while the second one
deals with evaluation and focusses on the product. We will apply these works
when analyzing our system in chapters 4 and 6. In order to keep this section
comprehensible for a general audience, we tried to reduce the formal exposi-
tions to the minimum necessary. On some parts, these become indispensible
to clarify concepts for those interested in technical detail, but they are still
accompanied with an informal description.

2.1.1. Characterizing Creativity in AI

Wiggins (Wiggins 2001; Wiggins 2003) views exploratory creativity as a

search for concepts in the space of all possible concepts of a domain (e.g.
in the music domain, it would be the set of all possible sequences of sounds).
This search is a priori constrained by a set of rules that define “acceptable”
elements (Boden’s conceptual space), R. This set could consist, for example,
of rules of style. This could comprise both creative and non-creative elements,
but all correct according to those rules of style. This set of elements is also
known (in (Wiggins 2001) as the set C .
Another set of rules that Wiggins proposes would be the set of “valued”
elements, those that, regardless of being “correct” or “according to style”,
become successful in solving the problem or achieve the goal that is being
sought (e.g. pieces of music that, although breaking stylistic rules, become
for some reason a good choice). This set, E , is much harder to define either
formally or informally, for often it is dependent on dynamic aspects (e.g.
aesthetic change, specific goals, specific context, personal mood, etc.).
Finally, the third set of rules proposed is the one that defines the strat-
egy followed when exploring the conceptual space, T . This represents the
choices, be they intentional or unconscious, that the creative agent (human
32 Creativity

or not) makes to travel in the space of possible elements. For example, some
people prefer to work “top-down”, i.e. to define the whole structure, then pro-
ceed reifying the ideas, others “bottom-up”. Others rely on ad-hoc method-
ologies, even randomness. It is not guaranteed, though, that all elements
found are acceptable (belong to C ). In other words, the application of T
may produce non-acceptable outcomes. This aspect is extremely important in
terms of creativity, as it opens a door to re-thinking the whole system.
According to Wiggins (Wiggins 2001), it is thus from the interaction of
these three sets, R, E and T that exploration of the universe of concepts
can be analyzed in the light of exploratory creativity. Each of these sets can
change with time, in the same way that science and arts evolve by chang-
ing their rules, goals or methodologies. It is this constant change that drove
Wiggins (and others (Wiggins 2001; Colton 2001; Ram et al. 1995)) to a con-
clusion that the process of exploratory creativity (the three sets) is also ex-
ploratory in and of itself. In other words, using the same formalization as the
one referred to above, Wiggins proposes the demonstration that transforma-
tional creativity is equivalent to exploration at the meta-level (Wiggins 2001).
The idea is that transformational creativity can only happen with changes in
(at least) one of the three sets and, if we jump one level up and consider
an exploration in the space of possible rule sets (e.g. the space of style rule
sets, space of “value judgments”, space of strategies), then the same analysis
can be made. Of course, the question arises: when to stop this recursion?; or
even, is this conclusive about the act of creativity? Aren’t those transforma-
tions driven bottom up or emergent (e.g. serendipity, empirical observation,
sensorial evolution), rather than a meta-level activity? Of course, Wiggins
setting is more a proposal for the analysis and discussion of creativity within
AI than an actual statement of how things work cognitively and, in that sense,
it has been an interesting base to apply.

2.1.2. Creativity assessment

While Wiggins centered his formalization on several aspects of the process

(the constraints, the strategy, the conceptual space), giving less attention to
the product, Ritchie proposes a set of measures11 for assessing creativity on
the basis of the results of the system (i.e. the product), its initial data and the
items that gave rise to its construction (the inspiring set) (Ritchie 2001).
Prior to describing the measures, we have to give a set of definitions. The
 Computational Creativity 33

first one regards the notion of basic item, an entity that a program produces.
“This is not a definition of what would count as successful or valid output for
the program, merely a statement of the data type it produces”. Ritchie pro-
poses two rating schemes to analyze the produced items: typicality ratings
(typ) and value ratings (val). There is also the notion of inspiring set, I, the
set of basic items that, explicitly or implicitly, lay behind a generative sys-
tem. For example, for a rule-based system, the elements implicitly defined by
those rules would be in the inspiring set; for a case-based system, the initially
stored cases would also be part of the inspiring set; for an example-based,
learning system, the examples given in the training phase would belong to
the inspiring set. The definition of the inspiring set is fundamental to estimate
how successful a system is in obtaining novel ideas, but it is often hard to find,
since the designer of the system is rarely conscious of all the influences be-
neath their own choice. For example, a system for composing a certain style
of music may have been designed via analysis of a set of pieces (which would
be part of its inspiring set), while if it were made from musicology theories,
defining that set would be a harder task.
Finally, we need to define four notation conventions. We assume that X
and Y are two sets of items, and that F can be a function such as typ or val:
de f
Tα ,β (X ) = {x ∈ X | α ≤ typ(x) ≤ β } : The subset of X falling in
a given range of normality
de f
Vα ,β (X ) = {x ∈ X | α ≤ val(x) ≤ β } : The subset of X falling in
a given range of quality
de f
AV (F, X ) = (∑x∈X F(x)/ | X |) : The average value of func-
tion F across finite set X
de f
ratio(X ,Y ) = | X | / | Y |: The relative sizes of two
finite sets X, Y

Ritchie proposes fourteen measures to assess the creativity of a system’s

output, R. Although it is assumed that R corresponds to the result(s) of a single
run, it is also suggested the generalization of these measures to a set of runs, in
order to cover the general behavior of the system. Thus, in this case, we invite
the reader to consider R as the set of results that a system has been able to
produce until a given point in time. The measures are intended to quantify the
behavior of the system in terms of average quality of results, their typicality
and of their ratios with regard to R and to the set of typical and valued items.
34 Creativity

In general, isolated measures will not say anything regarding creativity. Each
one contributes with its own part to the overall picture of the analysis of the
output. Some of them (the later ones), however, are combinations of the others
and allow us to conjecture about creativity potential.
measure 1 AV (typ, R) > θ , for suitable θ .
The first measure compares the average of typicality of items with a value,
θ . The following measure studies to what extent typical items form a signifi-
cant proportion of the results:
measure 2 ratio(Tα ,1 (R), R) > θ , for suitable α , θ .
measure 3 AV (val, R) > θ , for suitable θ .
measure 4 ratio(Vγ ,1 (R), R) > θ , for suitable γθ .
Measure 3 and 4 follow the same reasoning as the first two, but applied to
value (val). Alternatively, the fifth measure classifies the success of a system
as its ability to obtain a high proportion of highly valued items, within the set
of the typical ones:
measure 5 ratio(Vγ ,1 (R) ∩ Tα ,1(R), Tα ,1 (R)) > θ , for suitable α , γ , θ
The following three measures compare the set of highly valued, yet un-
typical results, to the output, to the whole set of untypical results and to the
set of typical highly valued outcomes.
measure 6 ratio(Vγ ,1 (R) ∩ T0,β (R), R) > θ , for suitable β , γ , θ .
measure 7 ratio(Vγ ,1 (R) ∩ T0,β (R), T0,β (R)) > θ , for suitable β , γ , θ .
measure 8 ratio(Vγ ,1 (R) ∩ T0,β (R),Vγ ,1 (R) ∩ Tα ,1 (R)) > θ , for
suitable β , γ , θ .
A program might be replicating its entire inspiring set, in which case it
might be said it is not creative since (at least some of) the outputs are not
novel:
measure 9 ratio(I∩R,I) > θ , for suitable θ .
Thus, a high value in this measure is something we don’t want if looking
for creativity in a system. Conversely, the system may produce outcomes that
do not belong to the inspiring set. Thus, in measure 10, we calculate the ratio
of all the items generated with respect to the ones (from that set of generated
ones) that also belong to the inspiring set:
measure 10 ratio(R,I∩R) > θ , for suitable θ .
Measures 11 and 12 propose some possible perspectives on the generated
items not belonging to the inspiring set: proportion of those that are typical
(measure 11); proportion of those that are valuable (measure 12). The former
stresses the capability of the system to produce results that are not replica-
tions, yet still fitting the “norms” (i.e. still being typical). The latter estimates
 Computational Creativity 35

how well the system is able to produce valued items that are not replications.
measure 11 AV (typ, (R−I)) > θ , for suitable θ .
measure 12 AV (val, (R−I)) > θ , for suitable θ .
Finally, measures 13 and 14 give an estimate about the proportion of
highly typical and valued novel results. The latter justly fits the view of cre-
ativity as the generation of “novel and valued” products.
measure 13 ratio(Tα ,1 (R−I), R) > θ , for suitable α , θ .
measure 14 ratio(Vγ ,1 (R−I), R) > θ , for suitable γ , θ .
These measures pose two obvious problems for their application. The first
one has to do with the rating schemes val and typ, namely the former would
demand a compromise that is rarely explicitly made in everyday observation
of creativity, of what a valuable outcome is composed of exactly. The second
problem regards the variables involved (α , β , γ and θ ). Finding acceptable
values will depend on experimentation in different contexts. Yet, until now,
there has been no application of these. Furthermore, their scales will differ
among measures (e.g. measures 4 till 9 yield values in the interval [0, 1],
measure 8 can give any positive real number, measure 10 always results in
values higher than 1). In this work, we assume α , β and γ to be 0.5.
Another issue is that Ritchie considers typicality and value, rather than
novelty and usefulness. While usefulness and value are often meant as syn-
onymous (in the sense that something is valued when it accomplishes a pur-
pose), typicality runs opposite to novelty. Assuming the risk of oversimpli-
fying these notions, we consider, in this work, that typicality is converse to
novelty (i.e. novelty(x) = 1−typ(x)) and value equals usefulness. This is im-
portant for the analyses made in chapter 6.

In terms of direct application to computational systems, there has been

almost no examples, and therefore empirical values and considerations for the
measures are not available. There may be three reasons for this happening:
the frameworks are still immature and therefore demand further work; there
has been no practical need for systems to measure their creative potential and
to compare with their peers, a fact that contrasts with the claim that some
make of being creative systems; they may be considered wrong or useless for
analyzing computational creativity. Regarding this latter possibility, it is clear
that these are the only accounts so far for the formal analysis of the creativity
of a computational system (except for even less developed formalizations,
such as the serendipity equations of Figueiredo and Campos (Figueiredo and
Campos 2001) or work on measuring surprise (Macedo and Cardoso 2001)).
36 Creativity

Since it is imperative to determine, even if not in an absolute fashion, the

creativity of the model we propose, we will apply these ideas to analyze our
system.
It is also patent that these two approaches have a lot in common. For exam-
ple, Wiggins R and E may correspond to Ritchie’s typ and val respectively.
There is however a fundamental difference: while Ritchie looks inside the
system’s results and their evaluative schemes and is dependent on them to
define every measure, henceforth considering only exploratory creativity12 ,
Wiggins’ formalism confronts the system with the universe (U ) of possible
items, thus allowing for meta-level analysis and therefore transformational
creativity. In figure 3, we summarize a classification of concepts within the
framework of Wiggins, also taking into account the inspiring set, the notions
of typicality and value and measures presented. This diagram is based on the
interaction between three sets:  R, T , E , the set of concepts that can
be obtained by the strategy T ; [[R]], the set of items defined by the rules R;
[[E ]], the set of valued items, acording to the rules E . Transformational cre-
ativity would thus have the effect of changing the set of reachable concepts.
The set “reachable creative concepts” has been named after the description
of the measures 5, 10, 11, 12, 13 and 14, but it can be argued that the label
“creative” should be given to the set of “valued untypical concepts”, since the
latter favours novelty and value.

2.2. Creative systems

2.2.1. What is a creative system?

Before proceeding, we must define the necessary (but not sufficient) condi-
tions for a computational system to be included in the list we study here as
being creative:
– It should aim to produce solutions that are not replications of previous
solutions (known to it).
– It should aim to produce solutions that are acceptable for the task it pro-
poses.
Of course, these correspond to the classical definition of novelty and use-
fulness, from a perspective of p-creativity, in Boden’s terms. Thus, we are
allowing for the classical extremes: a random process that can find good
items; a system that explores the whole search space with a brute-force blind
 Computational Creativity 37

Figure 3. The classification of concepts within the universe U according to the sets
 R, T , E , [[R]]and [[E ]]. Remember that C =[[R]]

method; a system that generates different outcomes every time, but which all
look very similar to humans. All these are creative systems, according to this
classification, but we intend primarily to rule out the (much larger) set of AI
systems that focus solely on the second condition: to accomplish a well de-
fined goal. As with intelligence, there is a continuum of creativity degree in
such systems, from the utterly non-creative to the undoubtedly creative, and
some dimensions can be inspected in order to sort out further classifications:
– The complexity it is able to treat, without breaking the conditions above
- creative abilities are often more salient (or vital) in problems with high
complexity. A program able to satisfy our criteria in highly complex set-
tings is definitely on the (highly) creative side of the continuum.
– Its ability to reason at different levels of abstraction - A system that can deal
with different levels of abstraction for the same problem would certainly be
closer to the goal of meta-level reasoning (and t-creativity) than a system
without such ability.
– Its ability to process different sorts of representation - It has been of-
ten referred in Creativity literature (e.g. (Karmiloff-Smith 1993)) that re-
38 Creativity

representation potentiates creative outcomes.

– Its ability to work in more than one domain - Creativity cannot be regarded
in isolation. A system that is able to deal with more than one domain with-
out structural changes should be considered more creative than one that
needs reprogramming to work with different domains. The former should
be closer to the ability of cross-domain transfer.
– Its ability to evaluate its own productions - Self-assessment or self-
criticism has also been mentioned as central to creative production. A sys-
tem that can evaluate its own productions would tend to be on the creative
side of the continuum.

2.2.2. Some systems and models

Of all the surveys presented in this book, the overview of the field of cre-
ative systems is the most difficult to make for two particular reasons: there
has been a surprisingly high proliferation of such systems during the past five
years, spreading across a variety of areas and approaches; only a few identify
themselves as “creative systems”, preferring different classifications such as
“cyber art”, “generative systems” or “creative design”. Moreover, from these,
only a few consciously follow an underlying “computational model of cre-
ativity”. Thus, in order to analyze the models that have been used so far, it is
necessary to abstract them from the existent implementations.
The list of systems chosen for the overview obeys three conditions. Each
system should have its implementation description available somewhere (rul-
ing out many commercial implementations), it should also have been pub-
lished recently13 (except for some classical examples), and it must satisfy,
even if only assumed informally, the definition above for creative systems.
The approaches presented range all over the traditional AI paradigm clas-
sification spectrum of Symbolic, Sub-symbolic and Hybrid systems. This di-
vision could therefore be a starting point to structure this overview. Neverthe-
less, we prefer to organize it according to issues that have been discussed so
far regarding the theme of creativity.
The first issue regards the opposition of perspectives cognition/society,
which were also observed in section 1. Some works follow the systems per-
spective of Csikszentmihalyi, via multi-agent environments in which creativ-
ity emerges as a result of multiple interaction. Examples of this approach are
the Hybrid Society (Romero Cardalda 1999), The Digital Clockwork Muse
 Computational Creativity 39

(TDCM) (Saunders and Gero 2001), Design Situations (DS) (Sosa and Gero
2003) and SC-EUNE (Macedo and Cardoso 2001). On the other side, there
are the systems that follow approaches based on a cognitive perspective (i.e.
the machine is one single agent) by applying domain-dependent rules (e.g.
Lothe’s Mozart Minuets (Lothe 2000), Aaron (Cohen 1981)), stochastic mod-
els (e.g. Craft and Cross’s fugal exposition generation (Craft and Cross 2003),
the Postmodernist generator (Bulhak 2000)), reusing past experience (e.g.
the Case-Based Reasoning approaches of ReBuilder (Gomes et al 2002) and
ASPERA (Gervás 2001)), evolutionary computation (e.g. NevAr (Machado
and Cardoso 2002) and Poevolve (Levy 2001)), modelling specific cogni-
tive phenomena like Metaphor (e.g. Sapper (Veale 1995; Veale and Keane
1997; Veale and O’Donoghue 2000)), Analogy (e.g. Copycat, (Hofstadter
and Mitchell 1988)) or Conceptual Combination (e.g. C3 (Costello, 1997)).
Of course, individual agents in multi-agent systems need to have individual
processes, and therefore these two perspectives are not totally incompatible.
A second issue regards evaluation: is the system performing self-
evaluation by any means in order to obtain the final product, or is there the
participation of a user, the generation of the product being a result of the
interaction (in this case, the general system - human plus machine - could
be seen as a two agent system)? In general, multi-agent systems presuppose
a built-in evaluation strategy, normally becoming part of the interaction be-
tween agents (agents reward other agents for positive evaluation, as happens
in TDCM), the Hybrid Society being an exception (humans can be part of
the egalitarian multi-agent system and produce and/or evaluate). In single-
agent architectures, some kind of self-assessment is also built-in, either via
probability, rules, or pre-trained mechanisms (like a Neural Network). These
self-assessments rely more on the appropriateness of the product according
to a style or goal than on aesthetic judgment. Yet, one should not read this too
strictly since we can find a variety of self-evaluation methods that consider
aesthetics (e.g. NevAr allows the application of aesthetic principles based on
perception; HR contains interestingness heuristics). Some systems rely on the
active participation of the user in the generation process. A typical example is
the interactive genetic algorithm (IGA), where the algorithm generates items
that are evaluated by the user (e.g. NevAr and (Sims, 1991)14 ), and this eval-
uation is used to produce the subsequent items. Max (Campos and Figueiredo
2001) is an agent that searches the Web for interesting pages (according to a
user profile) in order to trigger serendipitous insights for the user, who has to
give Max appropriate feedback to continue the cycle. MuzaCazUza and Re-
40 Creativity

Builder are two Case-based reasoning systems on music and software reuse
(respectively), which rely on the user for the adaptation of cases (in the case
of ReBuilder, it also provides an analogy mapping method of adaptation).
A third issue has to do with the use of memory. Does the system keep
track of past runs, and therefore is it able to profit from past experience? This
can be seen as an instantiation of the preparation phase discussed earlier. A
few systems have this property, namely SC-EUNE, NevAr, ReBuilder, Max,
Metacat and Sapper.
Being able to do meta-level reasoning, i.e. to reason about the method
of reasoning, could only be found in three systems: HR, Metacat and Sap-
per. The former (Colton, Bundy and Walsh 1999), named after Hardy and
Ramanujam, was designed by Simon Colton to carry out discovery in pure
mathematics. It performs a complete cycle of mathematics, including build-
ing its own mathematical concepts, making conjectures about the concepts,
attempting to prove the true conjectures and finding counterexamples for the
false ones. It builds new concepts according to seven heuristics and nine pro-
duction rules. All these are considered the most generic possibilities across
the field of mathematics. It has a meta-level reasoning version (Colton 2001),
in which it builds a high-level theory that contains concepts and conjectures
about the concepts and conjectures of the lower-level theory. For example, it
is able to form the high-level concept of “function” (there is a unique second
object for each first object found in the pairs which make up the examples
for these concepts). Examples of low-level functions are definitions for prime
numbers, perfect numbers, pairs and so on. The HR project has also a multi-
agent version with four HR-based agents with different strategies running in
parallel (Colton, Bundy and Walsh 2000), which cooperate by exchanging
new concepts and conjectures.
Metacat (Marshall 2002) is the latest evolution of Copycat (Hofstadter
and Mitchell 1988), which is a system for solving puzzle analogies (such as
“abc→abd::xyz→?”) that applies a bottom-up parallel strategy to find map-
pings between the source and the target, as well as explanatory relations
within them, and to associate these mappings and relations with concepts in
a Slipnet. This Slipnet is a semantic network with variable distance between
concepts (examples of concepts are the letters of the alphabet, relations like
opposite or predecessor, and attributes like rightmost or first letter of the al-
phabet) where a spreading activation algorithm is used to determine the plau-
sible mappings (e.g. “c” corresponds to “z” - both are rightmost - and “c” is
predecessor of “d” - in the alphabet). When a set of concepts is activated, a
 Computational Creativity 41

candidate solution is projected. If it fails, Copycat will slip to concepts that

have short distance to the activated ones (e.g. rightmost to leftmost) and try
other activation patterns. Metacat is a sibling of Copycat and it adds a set
of meta-level features. It keeps traces of the run and creates abstract themes,
which consist of pairs of Slipnet concepts. For example, a theme for repre-
senting the idea of alphabetic-position symmetry between “a” and “z” would
have the Slipnet concepts alphabetic-position and opposite. It keeps this new
information in an episodic memory that it uses to compare analogies between
different runs (it outputs texts like “this reminds me of the problem X...”). It
is also able to justify the reason behind a puzzle solution by the analysis of its
trace. However, it uses this meta-level knowledge (traces and themes) superfi-
cially in the sense that, unlike HR, it only uses it for communicative purposes
(with a user) and does not improve or change its own internal knowledge, or
even use previous solutions to solve present problems.
Sapper (which will be presented in detail in section 4.2) is also said to be
capable of meta-level modification, by altering the “construction space” that
guides the mapping process (Veale and O’Donoghue 2000). With this capa-
bility, the system should be able to propose different metaphor interpretations
for the same input data.
In spite of recurrently being asserted throughout the creativity theories in
section 1, the ability to do cross-domain transfer of ideas is absent in the ma-
jority of the systems. Even worse, the majority is tailored to work with its
own single domain. In this matter, again and surprisingly, HR presents itself
as the only one able to do cross-domain transfer (Steel 1999) and work with
different domains. However, these two capabilities rarely come together. The
cross-domain version of HR was built with special attention to concepts from
mathematical fields15 , while its application to domains separate from mathe-
matics (e.g. animal classification (Colton, Bundy and Walsh 2000)) has been
made in an isolated manner. Two other systems are capable of cross-domain
transfer, normally at an abstract level. Sapper is a metaphor interpretation
system that will be thoroughly presented in section 4.2, since it is impor-
tant for our work, while Drama will be briefly described in section 4.3. Both
find cross-space mappings between semantic networks, which can be used to
transfer knowledge, i.e. problem solving by analogy.
From this analysis, we may now abstract a taxonomy for computational
models of creativity:
– Systems model (SM). Creativity results from the interaction of a society
of agents. Each agent may have a similar or different role, and be imple-
42 Creativity

mented according to the methods below, but its interaction with others is
essential to find creative items.
– Evolutionary model (EM). Creativity emerges as a result of evolution of
the artifact. This evolution is made in parallel with concurrent streams of
candidate artifacts that eventually converge to a maximum. The judgments
(fitness functions) are either given by a user (IGA’s) or via algorithmic
methods such as neural networks or heuristic rules.
– Domain-centered model. Creativity results from expertise on a specific do-
main. Different domains invite different specific methods or knowledge
structures (even if the general approach remains the same). This model can
be divided into three sub-types:
– Expert Systems model (ESM). Items are generated by following well
established constraints and methodological rules of the domain of appli-
cation. Creativity is stimulated by allowing randomness in well-bounded
decision making points.
– Case-Based Reasoning model (CBRM). Creativity is the result of reuse
and adaptation of past experience with attention to the present context.
According to (Ram et al. 1995), this is achieved in five steps: problem
interpretation, problem reformulation, case and model retrieval, elabo-
ration and adaptation and evaluation.
– Stochastic model (STM). Creative items are generated from non-
deterministic automata that result from analysis on selected data16 .
When it is well-trained, issues like evaluation or memory are embedded
in the automaton, which rarely produce wrong outcomes or outcomes
that differ considerably from the initial data.
– Cognition-centered model (CCM). Creativity results from mental pro-
cesses that can be computationally modelled. It is domain-independent and
therefore items are represented at a conceptual level that needs to be reified
at application level. This reification may be made externally, but it must be
consistent with the concept description.
In tables 1 and 2, we give a summary of the characteristics of the systems
analyzed. Some systems are not described in detail here because their descrip-
tion would not add pertinent facts for this book. Many more systems were left
out (never mind the commercial ones), so this is a very small sample of the
state-of-the-art which hopes to cover the wide breadth of the approaches.
Perhaps due to the youth of the area of creative systems or to the differ-
ent purposes of each system, few provide formal analysis of the creativity
involved. The formalisms of Wiggins, Ritchie and the others have scarcely
Table 1. Generic description of the analyzed creative systems (part I)
Name Theme Reasoning Paradigm Domain
HR Concept Formation Rule-based Many (essentially Maths)
SC-EUNE Surprise Modelling Rule-based Architecture
TDCM Aesthetic evolution IGA+NN Artistic images
Design Situations Design Hybrid Design
Hybrid Society Aesthetic evolution IGA+NN Music
NevAr Image generation IGP Artistic images
MuzaCazUza Melody generation CBR Music
ReBuilder Software Reuse CBR+Analogy Software Engeneering
Max Serendipity Rule-based Web searching
Metacat Analogy Hybrid Puzzles
ASPERA Style imitation in poetry CBR Poetry
(Conklin and Witten 1995) Music prediction Stochastic Music
(Craft and Cross 2003) Style imitation in music Stochastic model (n-gram) Music
(Sims 1991) Image generation IGP Artistic images
Copycat Analogy Hybrid Puzzles
(Lothe 2000) Style imitation in music Rule-based Music (Mozart Minuets)
EMI Style imitation in music Stochastic model (n-gram) Music
(Manurung, Ritchie and Thompson 2000) Poem generation Hybrid (rule-nased+GA) Poetry
Poevolve Poem generation IGA+NN Poetry
WASP Poem generation Rule-based Poetry
JAPE Pun generation Rule-based Humour
Postmodernist generator Thesis generation Stochastic Literature/humor
MAKEBELIEVE Text Generation Rule-based Stories
Dupond Text Generation Rule-based Conversation
Brutus Text Generation Rule-based Stories
Aaron Image generation Rule-based Artistic images
HAHAcronym Acronym generation Rule-based Humour
Sapper Metaphor Hybrid Many
Drama Analogy Hybrid Many
Computational Creativity

C3 Conceptual Combination Rule-based Noun-noun compound

43
 44

Table 2. Generic description of the analyzed creative systems (part II)

Name Architecture Model Episodic Evaluation Reference
Memory
HR Allows both CCM, ESM No Built-in (Colton, Bundy and Walsh 1999)
Creativity

SC-EUNE Multi-agent CCM, SM Yes Built-in (Macedo and Cardoso 2001)

TDCM Multi-agent SM, EM No Built-in (Saunders and Gero 2001)
Design Situations Multi-agent SM No Built-in (Sosa and Gero 2003)
Hybrid Society Multi-agent SM No Interactive (Romero 1999)
NevAr Single agent EM Yes Inter./Built-in (Machado and Cardoso 2002)
MuzaCazUza Single agent CBRM No Interactive (Ribeiro et al. 2001)
ReBuilder Single agent CBRM, CCM Yes Interactive (Gomes et al 2002)
Max Single agent CCM, ESM Yes Interactive (Campos and Figueiredo 2001)
Metacat Single agent CCM Yes Built-in (Marshall 2002)
ASPERA Single agent CBRM No Built-in (Gervás 2001)
(Conklin and Witten 1995) Single agent STM No Built-in
(Craft and Cross 2003) Single agent STM No Built-in
(Sims 1991) Single agent EM No Interactive
Copycat Single agent CCM No Built-in (Hofstadter and Mitchell 1988)
(Lothe 2000) Single agent ESM No Built-in
EMI Single agent STM No Built-in (Cope 1991)
(Manurung, Ritchie and Thompson 2000) Single agent EM,ESM No Built-in
Poevolve Single agent EM No Built-in (Levy 2001)
WASP Single agent ESM No Built-in (Gervás 2000)
JAPE Single agent ESM No Built-in (Binsted 1996)
Postmodernist generator Single agent STM No Built-in (Bulhak 1996)
MAKEBELIEVE Single agent ESM No Built-in (Liu and Singh 2002)
Dupond Single agent ESM No Built-in (Mendes 2004)
Brutus Single agent ESM No Built-in (Bringsjord and Ferrucci 2000)
Aaron Single agent ESM No Built-in (Cohen 1981)
HAHAcronym Single agent ESM No Built-in (Stock and Strapparava 2003)
Sapper Single agent CCM Yes Built-in (Veale and Keane 1997)
Drama Single agent CCM No Built-in (Eliasmith and Thagard 2001)
C3 Single agent CCM No Built-in (Costello, 1997)
 Computational Creativity 45

been applied, which is understandable given the problems which arise when
determining the inspiring set or the value for the many variables involved.
Only WASP (Gervás 2002) and Dupond (Mendes 2004) have been analyzed
so far with those formalisms, and it is still complicated to compare them
with other systems (as emphasized in (Pereira, Mendes, Gervás and Cardoso
2005)). In this book, we give another contribution to this fundamental aspect
of evaluation of creative systems.
We would like to conclude this section by noting that, in spite of the cur-
rent proliferation of creative systems, the large majority is exploratory. One
can say that transformational creativity has so far been achieved by systems
such as HR and Metacat, although only at an elementary level. These systems
deal with meta-knowledge but are still far from actually transforming their
conceptual space, strategy, knowledge representation or evaluation function.
Chapter 3
Working with Concepts

Since in this book we propose a model of concept invention, it is therefore

imperative to define what exactly is meant by a concept and by concept in-
vention. Furthermore, we have to present and explain in detail the cognitive
and computational basis applied at the level of working with those concepts.
Thus, in this chapter, we define Concept and Concept Invention (and oppose
it to concept formation). We also introduce Conceptual Combination, Con-
ceptual Blending, Metaphor and Analogy, which occur in different parts of
this book. All these work at the level of concepts or networks of concepts.

1. What is a concept?

Perhaps the most specific definition we can give is that a concept is an ab-
straction that refers to ideas, objects or actions. Concepts can be dynamic en-
tities, i.e. they can change with time (e.g. the concept of “phone” has evolved
along with its technology), person (e.g. for some people a “crocodile” is a
“pet”, while for others it is not) or context (e.g. the concept of “giant” will
differ radically when comparing an “elephant” with a “human” and with a
“dinosaur”). In some domains, normally scientific, they can also be formal
and static (e.g. the concept of “prime number” is not supposed to change).
More than about the definition, much debate has been about how concepts
are represented in cognition. There are three main views:
– Prototype view (Rosch 1975). Concepts are represented in the mind by
prototypes, rather than by explicit definitions, which can be used to differ-
entiate when an instance is or is not an example of the concept. Concepts
are represented by an “idealized” prototype, which has the “average” char-
acteristics of the concept (e.g. the prototype of “bird” would have “has
wings”, “has feathers”, etc.) or by a “paradigmatic” prototype (e.g. a cof-
fee cup for a cup or wooden spoon for large spoon). Of course, this view
raises problems because concepts are not necessarily static entities, defin-
able with a fixed set of properties.
– Exemplar view [?]. Concepts are represented by their most common ex-
emplars. Therefore, classifying an instance consists in determining which
48 Working with Concepts

remembered exemplars are the most similar. This view implies that we
organize experience in an episodic memory. If considered in isolation, the
exemplar view fails in a particular aspect. Although it is agreed that knowl-
edge is dependent on individual experiences, the ability to do abstraction,
to generalize from experience, is fundamental, otherwise memory would
be insufficient for reasoning.
– Theory view (Murphy and Medin, 1985). The representation of concepts
is based on micro-theory. A micro-theory describes the concept with facts
about the concept (or related concepts) and causal connections between
them. For example, the concept “bird” would have the facts that “it flies”,
“it has wings“, etc., but also rules that explain causality (e.g. Why do birds
fly? Why do they nest in trees?). Thus, a micro-theory can be seen as com-
prising a concept network (with causal links) and rules about the concept.
This view also poses some problems such as these two: what should the
limits of a micro-theory be (e.g. should we explain flight by physical rules,
or with common sense and to which level of detail)? Since concepts can be
dynamic, representing them with a theory would raise all sorts of problems
of non-monotonic reasoning (how to represent change? how to maintain
consistency and tractability?).
In AI, these three views have been applied. To name a few examples:
the prototype view is common in systems that represent concepts as attribute
value sets, such as in some machine learning systems (e.g. version space
learning (Mitchell 1978), decision trees); the exemplar view is typical in
Case-Based Reasoning systems, where episodic memory is used to compare
old to new problems; the Theory view is common in Logics (for example in
Inductive Logic Programming) and in systems that use semantic networks,
such as Sapper and Copycat (presented in sections 4.2 and 2.2.2).
Throughout this book, whenever we refer to concepts, we assume the The-
ory view, both in relation to our work and to the work of others, except when
explicitly stating an alternative. It is also important to state the relationship
between concept and category. In our work, a category is itself also a con-
cept, but viewed from the perspective of membership (e.g. the concepts dog
and wolf belong to the canine category, while the concepts canine and feline
belong to the mammal category). This follows an AI Ontology Engineering
philosophy, in which an ontology contains a web of interrelated concepts,
sometimes gathered within more compreehensive concepts (the categories,
which are also concepts themselves).
 Building concepts 49

2. Building concepts

Throughout the literature, there seems to be some confusion with the no-
tions of concept discovery, formation, invention, generation, design and cre-
ation. Sometimes they are synonymous to each other, sometimes they are
considered different. We propose a distinction between two ways of build-
ing concepts: concept formation and concept invention. We provide a con-
sensual definition for concept formation, from Psychology, which coincides
(also consensually) with concept discovery. The definition for concept inven-
tion (or generation or creation) may be less agreed upon since it is based on
less formal principles.

2.1. Concept Formation

In Psychology, Concept Formation (also known as concept learning or con-

cept discovery) is associated with the development of the ability to respond
to common features of categories of objects or events. In forming a concept,
the subject must focus on the relevant features and ignore those that are ir-
relevant. In AI, this task is normally taken by machine learning, in which
patterns are abstracted from analysis of data. In fact, the goal of machine
learning systems is to form concepts. In this sense, if these systems happen
to favor deliberately the formation of novel concepts (as opposed to systems
built for well-defined goals, such as a pattern detection Artificial Neural Net-
work, or a decision tree for classification), they can be classified as creative
systems. A good example of this is the scientific discovery field. Again, we
refer to HR (Colton, Bundy and Walsh 1999) since we give for it is the most
recent one of a series of systems centered on mathematics concept formation
(e.g. AM and EURISKO (Lenat 1984)) that use machine learning.
We can see from these definitions that concept formation is more con-
cerned with analysis than with synthesis. In other words, works about con-
cept formation deal more with abstraction of data regularities than with the
invention of novel concepts17 . There is a fundamental difference between
these two processes: the former is based on finding sound explanations for
data regularities, while the latter on producing concepts without concerns of
soundness. However, this does not necessarily mean that concept formation
is not creative. Quite the opposite, the capacity to perceive regularities and
associations that no one has found before is definitely behind many of the
major achievements of humanity.
50 Working with Concepts

2.2. Concept Invention

A concept has been invented (as opposed to formed) when it cannot be de-
duced from its generative process and when it did not exist before, inten-
sionally or extensionally. In other words, the process of concept invention
is unsound. This definition covers a very broad range of possibilities, from
randomness to heuristics-based search.
In concept invention, evaluation becomes a fundamental issue. Since there
is no a priori notion of validity, criteria must be met for the assessment of
the generated concepts. These criteria can coincide with those discussed for
creativity (i.e. novelty and usefulness) or to problem solving (i.e. satisfying
a goal). Since these are, again, difficult criteria, concept invention is nor-
mally applied as a generative phase to feed other sound procedures, which
can guarantee validity. For example, in scientific discovery systems (e.g. HR,
AM, EURISKO), conjectures are generated from the application of heuris-
tics; in conceptual design the process of concept invention (or generation)
commences by establishing structural relationships and searching for regular-
ities and combining them into concept variants (Reffat 2002). In AI systems
in general, concept invention has been implemented based on heuristics (e.g.
in HR), parallel processes (e.g. in Copycat), evolutionary techniques (e.g. in
NevAr), to name a few. The main argument here is that in neither case is
the novel concept the logical conclusion from data analysis, but merely a
bounded guess to be explored later.
To conclude, we must stress that, in practice, there is not such a strict sep-
aration between formation and invention (rather, there is a continuum). Every
discovery involves conjecturing (i.e. inventing - or speculating about - new
concepts yet to be proven), a process that has a great deal of its power in un-
sound processes, like aesthetics, intuition and free-association. In the systems
referred to above (HR, AM and EURISKO), the conjecture generation step is
fundamental and it is achieved with the application of production rules and
heuristics to evaluate how interesting yet-to-be-proven concepts are.
The distinction between formation and invention could be reduced to a
problem of constraint satisfaction: formation has stronger constraints to sat-
isfy than invention, which is ill-defined. However, a clearer distinction is
needed since these correspond to two distinct, yet inter-dependent, steps of
creativity: rationality and imagination. Once again, convergence and diver-
gence. While rationality is more constrained, thus more limited but compu-
tationally implementable, imagination allows a world of possibilities that for
 Mixing concepts 51

a formal machine are hardly feasible. We propose that, at least philosophi-

cally, we should consider these two forms of working with concepts when
attempting to model computational creativity.

3. Mixing concepts

3.1. Conceptual Combination

Conceptual combination is the process of combining two or more concepts

together, often resulting in a novel concept with an emergent structure of
its own. Although regarded as a universal cognitive process, the main mo-
tif of study in the conceptual combination community is language composi-
tionality, more specifically interpretation of noun noun (e.g. “pet fish”) and
adjective noun (e.g. “blue cup”) combinations. In this context, the first word
(“pet”, “blue”) is called modifier, while the second is the head (“fish”, “cup”).
Four types of combination are proposed (see e.g. (Hampton 1997; Keane
and Costello 2001)): relational, property mapping, conjunctive and known-
concept. Relational combinations establish some relationship between the
modifier and the head (e.g. in “bed pencil”, “a pencil that you put beside
your bed for writing some messages” (Keane and Costello 2001)); property
mapping involves a property of one concept being asserted to the other (e.g. in
“bed pencil”, “a pencil shaped like a bed” (Keane and Costello 2001)); con-
junctive combinations conjoin both concepts in some way, the interpretation
being both the modifier-concept and the head-concept (e.g. in “bed pencil”,
“a big, flat pencil that is a bed for a doll”); known-concepts or lexicalized
compounds are those that are commonly used and established in communica-
tion (e.g. “pencil case”), sometimes effectively forming a single lexical unit
(e.g. “railway” or “lipstick”).
There are four main theories for conceptual combination: Abstract rela-
tions (Gagné and Shoben, 1997); Dual-Process (Wisniewski, 1997); Com-
posite Prototype (Hampton 1987); and Constraints (Costello 1997). The ab-
stract relations theory says that only a limited number of predicates can relate
the modifier with the head noun: CAUSE, HAS, MAKES, MADE OF, FOR,
IS, USES, LOCATED, DERIVED FROM, ABOUT, DURING, and BY. The
dual-process theory proposes two kinds of processes for conceptual combi-
nation, structural alignment and scenario construction. Structural alignment
explains property and conjunctive interpretations (there is an alignment of
52 Working with Concepts

attributes of both concepts), while scenario construction explains relational

interpretations (e.g., “a night flight is a flight taken at night”).
The other two theories are important for our work, so we will pay them
more attention. The composite prototype model of James Hampton focusses
on conjunctive combinations (e.g. “pet bird”) and proposes that, when form-
ing a concept such as “pets” that are also “birds”, people take their prototype
representations of “pet” and “bird” and combine the prototypes into a com-
posite to represent the conjunction. This new concept then inherits its own
attribute values from one or the other constituent parent according to certain
principles. For example, the location slot for “pet” has the value in the home,
while the same slot for “bird” has the value in the wild. “Pet bird” inherits
the value from “pet” rather than from “bird”. On other attributes, the oppo-
site might happen (e.g. the slot covering: feathered - is inherited from “bird”
rather than from “pet”, where its most common value is furry) (Hampton,
1997). Two influences on attribute inheritance are: the centrality of the at-
tribute (e.g. the location of “pet”, as staying at home, is central, while color
is not); its possibility in the composite (e.g. the value migrates could not be
possible in a “pet bird”). Another important aspect of the composite proto-
type model is emergence: attributes which are considered true of the con-
junction, but not true of either constituent. It appears that a major source of
emergent properties is simply knowledge of the world - or “extensional feed-
back” (Hampton 1987). Although the author argues that “we can not expect
any model of conceptual combination to account directly for such effects”
(Hampton 1997), he presents two specific sources of emergence: exemplar-
based, in which typicality of items in conjunctive categories can vary as a
function of the kinds of exemplar found in those categories (e.g. a “small
spoon” is typically made of metal, while a “large spoon” is not); theory-based,
a background theory is applied to infer emergent attributes (e.g. a “beach bi-
cycle” must have particularly wide tyres). All these ideas from James Hamp-
ton are also explored in Conceptual Blending, presented in section 3.2.
The Constraints theory of Fintan Costello and Mark Keane (Costello
1997) describes conceptual combination as a process which constructs repre-
sentations that satisfy the three constraints of diagnosticity, plausibility and
informativeness. Diagnosticity requires the presence of diagnostic properties
from each of the concepts being combined. The diagnostic properties of a
concept are those which occur often in instances of that concept and rarely
in instances of other concepts (similar to salience (Milosavljevic and Dale
1996)). The plausibility constraint requires the preference to semantic ele-
 Mixing concepts 53

ments which are already known to co-occur on the basis of past experience.
This constraint would predict that the interpretation “an angel pig is a pig with
wings on its torso” would be preferable to “an angel pig is a pig with wings
on its tail”. Informativeness requires an interpretation to convey a requisite
amount of new information. Informativeness excludes feasible interpretations
that do not communicate anything new relative to either constituent concept;
for example, “a pencil bed is a bed made of wood” (Keane and Costello 2001).
Costello and Keane implemented a computational model of their theory.
The system is named Constraints on Conceptual Combination (or C3 ) and
will be subject to a comparison with Divago in chapter 5.

As could be seen, conceptual combination is viewed as a cognitive pro-

cess, although its analysis is usually constrained to a particular language,
primarily to English. There are, however, aspects which are specific to some
language or family of languages. Indeed, while Dutch and German also al-
low the same kinds of combinations, other languages such as Portuguese or
French don’t. In Portuguese, the use of prepositions (e.g. de, para, etc.) in
combinations guarantees non-ambiguity. In principle, two consecutive nouns
in Portuguese correspond to a conjunctive interpretation (and, rarely, to a
property interpretation). Of the four theories presented, only the composite
prototype model seems to be totally language independent, perhaps because
it is directed to conjunctive combinations. This does not mean that the other
theories are less valid or unrelated to cognition or creativity, rather we say
that they favor problems at the level of language rather than at the level of
concepts. In this sense, they are models of interpretation and natural lan-
guage understanding, without paying attention to other domains like visual
arts, music or scientific discovery. In these domains too, conceptual combi-
nation is constantly present, and is often shared across different fields and
media (e.g. in the Baroque style, abstract concepts such as ornamentation or
luxury, travel across the several fields). Thus, we conclude that these mod-
els are a priori limited as models of creativity (except for Hampton’s model,
since it focusses on generic concepts and considers emergence).

3.2. Conceptual Blending

The framework of Conceptual Blending (CB), also known as Conceptual In-

tegration, was developed by Gilles Fauconnier and Mark Turner, and was
54 Working with Concepts

Generic Space

Input 1 Input 2

Blend

Figure 4. The original, four-space model, of Conceptual Blending

initially motivated towards specific cognitive phenomena such as Metaphor,

Metonymy and Counterfactual Reasoning (Fauconnier and Turner 1998).
Blending is generally described as involving two input knowledge structures
(the mental spaces) that, according to a given mapping, will generate a third
one, called Blend. This new domain will maintain partial structure from the
inputs and add an emergent structure of its own.
More recently Fauconnier and Turner propose CB as an explanation for
various cognitive phenomena. This claim has been taken even further in the
book “The way we think”(Fauconnier and Turner 2002), in which the authors
suggest their framework for explaining “the nature and origin of cognitively
modern human beings”. These bold claims incurred the voice of the critics,
which we will summarise once we have explained CB in some detail. As the
reader will see, we agree with these critics but we think that some of the
essential components of Conceptual Blending should be given merit. Most of
all, we think that CB as a framework, and in the way it is presented (mainly
 Mixing concepts 55

before (Fauconnier and Turner 2002)), allows it to be considered by AI, and

more specifically by Computational Creativity. The question now arises of
whether it could be considered per se a model of creativity. According to
our views given above (in chapter 2), we can hardly say that CB, alone, is a
model of creativity: indeed, there is nothing deliberately dedicated to novelty
in the framework; there is no proposed method for distinguishing whether a
blend is or is not creative (e.g. “red pencil” is considered to have complex
structure - does this make “red pencil” creative?); there is no clue for how or
why a pair of inputs should be chosen to potentiate creativity. On the other
hand, it proposes a many-step method for bisociation (that indeed follows a
conceptual combination philosophy, such as proposed by James Hampton,
briefly described above (Hampton 1987), which we consider as vital for the
Creativity Model presented in this book. And, although formally undefined
in so many issues, it is more detailed and explained than the other models (of
combination) studied. Thus, it is the combination of CB with an AI search
method, together with novelty and usefulness heuristics, that we will propose
as a creative system.
As just said, Conceptual Blending is not formally or algorithmically de-
scribed in its fundamental details by Fauconnier and Turner. As far as we
know, there are only a few formal accounts of this subject, apart from our
own (Goguen 1999; Veale and O’Donoghue 2000; Lee and Barnden 2001)
and its formalization is clearly not given priority within the mainstream CB
community (see discussion about computational modelling in (Fauconnier
and Turner 2002: 110-111).
The computational realization of this model is definitely a big challenge
since Conceptual Blending has many particularities that vary according to the
situation, complex components like intuition, social behavior, expectation or
common sense. In other words, there are several issues that are clearly hard
to model. Yet, the intersection of AI and CB may bring, if not the computa-
tional model of the framework, at least methods or algorithms that may bring
important contributions for the field of Computational Creativity, as we hope
to demonstrate.

3.2.1. The Framework

A blend is a concept or web of concepts whose existence and identity, al-

though attached to the pieces of knowledge that participated in its genera-
56 Working with Concepts

tion (the inputs), acquires gradual independence through use. We often find
a blend as being a concept that has the structure of other concepts, yet also
having its own (emergent) structure. We find examples of blends in many
sorts of situations. People have been making blends from at least the times
of Greek mythology (e.g. Pegasus) till today (e.g. the Pokemon creatures).
They are present throughout our daily communication (e.g. “John digested
the book”), technological evolution (e.g. “Computer virus”, “Computer desk-
top”), arts (e.g. Mussorgsky’s “Pictures at an exhibition”; Kandynsky’s “Im-
provisations”), advertising (e.g. Swatch is a blend of “swiss” and “watch”).
The works of (Mandelblit 1997; Sweetser and Dancygier 2005; Coulson
2000) and (Veale and O’Donoghue 2000) contain or are in themselves exam-
ples of how CB can contribute to Linguistics, Creative Cognition, Analogy
and Metaphor.
The first fundamental element of Conceptual Blending is the mental space.
A mental space is “a partial and temporary representational structure which
speakers construct when thinking or talking about a perceived, imagined,
past, present or future situation” (Grady, Oakley and Coulson, 1999). “Men-
tal spaces are small conceptual packets constructed as we think and talk, for
purposes of local understanding and action. (..) [they] are very partial. They
contain elements and are typically structured by frames. They are intercon-
nected, and can be modified as thought and discourse unfold. Mental spaces
can be used generally to model dynamic mappings in thought and language”
(Fauconnier and Turner 2002: 40-41). From a symbolic AI perspective, a
mental space could be represented as a semantic network, a graph in which
we have nodes identifying concepts18 (corresponding to the elements of a
mental space) interconnected by relations. The definitions of mental space
still allow many other representations (e.g. cases in Case-Based Reasoning,
memes in Memetics or even the activation pattern of a Neural Network in a
given moment) but these would certainly demand more complex computa-
tional treatment, especially with regard to the mapping. In figure 5, we show
two possible mental space representations for “computer” and “virus”.
In order to generate a blend, we must find mappings between the two
mental spaces. We call these cross-space (or cross-domain) mappings. They
connect elements of one mental space to others, in another mental space. A
mapping may be achieved through different processes (e.g. identity, structure
alignment, slot-filling, analogy) and doesn’t have to be 1-to-1, i.e., an element
may have more than one counterpart or it can have no counterparts at all. A
possible mapping for the “computer virus” blend is shown in figure 6.
 Mixing concepts 57

computer virus
binary replicates

computer ability
virus
representation host
contain invades
program infects
processes reduced
collection use
belong_to
instruction produce resources
capacity
receive evaluative_quality
output

input unwanted

Figure 5. Two simple mental spaces for computer program and for virus

computer virus
binary replicates

computer ability
virus
representation host
contain invades
program infects
processes reduced
collection use
belong_to
instruction produce resources
capacity
receive evaluative_quality
output

input unwanted

Figure 6. Cross-space mapping between the mental spaces of computer and virus

Another important notion is that of frames. When “elements and relations

are organized as a package that we already know about, we say that the men-
tal space is framed and we call that organization a frame” (Fauconnier and
Turner 2002: 102-103). A frame is therefore a kind of abstract prototype
of entities, actions or reasonings. For example, the mental space of “bus”
could be organized according to the frame “transport means”, while the men-
tal space of “Mary’s wedding” could be organized by the “marriage” frame.
Of course, this reminds us of the frames and scripts from early AI research
but, in this case, these frames may be dynamic (they can change with time,
individual and context, as with concepts in general) as well as compositional
(there are many layers of abstraction for frames). We call the principal frame
58 Working with Concepts

Virus Frame
virus
binary computer replicates

computer ability
virus
representation host
contain invades
program infects
processes reduced
collection use
belong_to
instruction produce capacity resources
receive evaluative_quality
output

input unwanted

Program Frame

Figure 7. The organizing frames of the mental spaces of computer and virus

underlying a mental space organizing (e.g. “transport means” is the organiz-

ing frame of “bus”, while “container” is not). In figure 7, we present two
organizing frames of the “computer” and “virus” mental spaces (the program
frame and the virus frame).
The frames preserve order in the blend, in the sense that they guide the
process of blend construction to recognizable wholes. This does not mean,
however, that the blend will integrate one single frame. Sometimes, as in the
example of “computer virus” (in figure 8), the blend will inherit structure
from both frames. As we can see in this example, some elements may not be
projected. In this case, the “input” and “output” elements of computer viruses
are normally a lot more subtle (or hidden) than in the usual programs.
In CB, the generation of a blend takes three (not necessarily sequential)
steps:
– Composition.
“Projection of content from each of the inputs into the blended space.
Sometimes this process involves the ’fusion’ of elements from the inputs
(...)” (Grady, Oakley and Coulson, 1999). Taken together, the projections
from the inputs make available new relations that did not exist in the sep-
arate inputs. The paired elements are projected onto the blend as well as
other surrounding elements and relations. This is a selective projection,
i.e., some elements get projected to the blend, some don’t.
– Completion. “The filling out of a pattern in the blend, evoked when struc-
ture projected from the input spaces matches information in long-term
 Mixing concepts 59

computer virus
binary replicates

computer ability
virus
representation host
contain invades
program infects
processes reduced
collection use
belong_to resources
instruction produce capacity
receive evaluative_quality
output

input unwanted

binary replicates

ability
representation
collection
instruction
contain
processes
invades
computer infects
computer virus

reduced

use
capacity belong_to

evaluative_quality
resources

unwanted

Computer Virus

Figure 8. The computer virus blend

memory. In this way, the completion process is often a source of emer-

gent content in the blend” (Grady, Oakley and Coulson, 1999). Knowledge
of background frames, cognitive and cultural models, allows the composite
structure projected into the blend from the inputs to be viewed as part of
a larger self-contained structure in the blend. The pattern in the blend trig-
gered by the inherited structure is “completed” into the larger, emergent
structure.
– Elaboration. “The simulated mental performance of the event in the blend,
which we may continue indefinitely” (Grady, Oakley and Coulson, 1999).
60 Working with Concepts

The structure in the blend can then be elaborated. This is called “running
the blend”. It consists of cognitive work performed within the blend, ac-
cording to its own emergent logic.
We illustrate the process of blending construction with another classic
example, the “Riddle of the Buddhist Monk”, which comes from Arthur
Koestler’s The Act of Creation:

A Buddhist Monk begins at dawn one day walking up a mountain, reaches

the top at sunset, meditates at the top for several days until one dawn when he
begins to walk back to the foot of the mountain, which he reaches at sunset.
Make no assumptions about his starting or stopping or about his pace during
the trips. Riddle: Is there a place on the path that the monk occupies at the
same hour of the day on the two separate journeys?

Following Koestler’s solution to the riddle, Fauconnier and Turner suggest

that it involves blending two different input spaces, one concerning the up-
ward trip (day ‘d1 ’) and another one concerning the downward trip (day ‘d2 ’).
A generic space holds the commonalities between the two input spaces (a
moving individual, his position, a path linking foot and summit of the moun-
tain, a day of travel, and motion in an unspecified direction) (Fauconnier and
Turner 2002: 45-46). The reasoning thus goes: first, composition of elements
from the inputs makes relations available in the blend that do not exist in
separate inputs. Only in the blend we have two individuals instead of one.
Second, completion brings additional structure to the blend (e.g. the frame
of “two people starting a journey at the same time from opposite ends of a
path”). At this point, the blend is integrated: it is an instantiation of a familiar
frame (“two people starting a journey...”). By virtue of this frame, we can now
run the blend, i.e., elaborate it. In this case, it coincides to applying intuitive
movement laws in opposite directions, which will make the two imagined
monks, a1 and a2 , eventually meet each other at some point, thus answering
the riddle. This is what Fauconnier and Turner call emergent behavior. This
reasoning is schematized in figure 9 (Fauconnier and Turner 2002: 43-44).
This example also leads us back to the discussion of Conceptual Blending
and Creativity above. CB does not actually tells us to blend the input spaces
so that the monk performs both journeys on the same day. This is a crucial
step to distiguish a creative from a non-creative solution (or a non-solution)
to the puzzle. Rather, once we make this decision, CB allows us to perform
the blend and determine the results (i.e., that a point does exist). In fact, there
is an initial step of high complexity (which inputs to choose to solve a prob-
 Mixing concepts 61

Generic
Space

a2
d1 d2
a1
Input
Space 2
Input
Space 1
d'

Blended
Space

Figure 9. Blended spaces for the Riddle of the Buddhist Monk

lem?) that is so far not accounted by the CB framework.

In general, the projection of data to the blend is not exhaustive or predeter-
mined in any way. Each element may be projected “untouched” to the blend,
it may be “fused” with other concepts, it may be projected to another element
(usually the projection of its counterpart) and it may not even be projected.
This selective projection brings considerable complexity to the blending pro-
cess because it raises the number of possible combinations to an extremely
large number. This space of potential blends will certainly not be completely
traversed for a new blend construction. In a quite different manner, projec-
tions are selected through a constraint-guided process of accommodation to-
wards satisfying a set of Optimality Principles (Fauconnier and Turner 2002):
– Integration - The blend must constitute a tightly integrated scene that can
be manipulated as a unit.
– Pattern Completion - Other things being equal, complete elements in the
blend by using existing integrated patterns as additional inputs.
– Topology - Other things being equal, for any input space and any element
in that space projected into the blend, it should be optimal for the relations
of the element in the blend to match the relations of its counterparts.
62 Working with Concepts

– Maximization of Vital Relations - Other things being equal, maximize the

vital relations in the network. Turner and Fauconnier identify 15 differ-
ent vital relations: change, identity, time, space, cause-effect, part-whole,
representation, role, analogy, disanalogy, property, similarity, category, in-
tentionality and uniqueness.
– Intensification of Vital Relations - Other things being equal, intensify vital
relations.
– Web - Manipulating the blend as a unit must maintain the web of appropri-
ate connections to the input spaces easily and without additional surveil-
lance or computation.
– Unpacking - The blend alone must enable the perceiver to unpack the blend
to reconstruct the inputs, the cross-space mapping, the generic space, and
the network of connections between all these spaces
– Relevance - Other things being equal, an element in the blend should have
relevance, including relevance for establishing links to other spaces and for
running the blend. I.e. it should have a good reason to exist.
These constraints work as competing pressures and their individual influ-
ence in the process should vary according to the situation; when the value
of one grows, others may decrease. As far as we know, there is no work yet
towards an objective study of the optimality pressures, measuring examples
of blends or formally specifying these principles. This, we believe, inhibits
considerably the appreciation and application of Conceptual Blending in sci-
entific research, thus a lateral motivation for the work presented here becomes
that of testing and specifying a formal proposal of these optimality pressures.
Among the many possible classifications of blending networks, Faucon-
nier and Turner particularly stress three kinds: mirror networks, single-scope
and double-scope networks. A mirror network is one in which both input
spaces share the same organizing frame, and so does the blend. In single-
and double-scope networks, input spaces have different organizing frames.
In a single-scope network, the blend has the organizing frame of only one of
the input spaces, while, in a double-scope network, the organizing frame of
the blend results from a combination of the inputs. The latter, as argued by
(Fauconnier and Turner 2002), is fundamental for human modern cognition
and is deemed more creative. In the examples above, “computer virus” can be
considered a double-scope blend, since its structure comes from a combina-
tion of the input’s structures, while the “Buddhist monk” is a mirror network
because both input spaces share the exact same structure (both have a path, a
monk, a direction, a mountain, etc.), the only difference being the direction.
 Mixing concepts 63

CB is a promising contribution to modelling creativity since it consists

of the generation of new concepts from integration of previous knowledge,
with its own emergent structure (the whole is bigger than the sum of its
parts); it has a domain independent evaluative methodology (e.g. the optimal-
ity constraints); it is consistent with processes often associated with creativ-
ity such as Metaphor and Conceptual Combination (Fauconnier and Turner
1998; Veale and O’Donoghue 2000; Coulson, 2000); and, finally, it is not,
by itself, a deterministic process, rather it is a context-sensitive process that
considers “in parallel” a set of constraints that may interact and yield a (po-
tentially large) varied space of equally valid solutions. In other words, with
the same starting conditions, we may get many different results, all with the
same overall value.
Since, in this book, we propose a computational model of Conceptual
Blending, we will also apply some established examples. We call these es-
tablished because they have often been cited in various articles or represent
classic situations approached in Blending literature. They were taken from
the main literature references (namely from (Fauconnier and Turner 2002)
and (Coulson 2000)). In section 6, we will apply them for purposes of vali-
dation of our blending module. All the examples are described in Appendix
B.
So far, there is little work on computational Conceptual Blending. We can
only name three examples: Joseph Goguen’s Algebraic Semiotics’ approach
(Goguen 1999); Veale and O’Donoghue’s extension of Sapper to account for
Blending (Veale and O’Donoghue 2000); and Barnden and Lee’s approach to
Conceptual Blending with counterfactuals (Lee and Barnden 2001). Joseph
Goguen proposes to describe blending using the algebraic semiotics formal-
ism. Algebraic semiotics is a formal theory of complex signs addressing in-
terface issues, in a general sense of “interface” that includes user interface
design, natural language and art. In the context of Semiotics, there are two
views of what a sign is, from Saussure (Saussure 1983) and Peirce (Peirce
1958). Goguen follows mainly Peirce’s definition19 :

A sign (..) [in the form of a representamen] is something which stands to

somebody in some respect or capacity. It addresses somebody, that is, creates
in the mind of that person an equivalent sign, or perhaps a more developed
sign. That sign which it creates I call the interpretant of the first sign. The sign
stands for something, its object. It stands for that object, not in all respects,
but in reference to a sort of idea, which I have sometimes called the ground
of the representamen. (Peirce 1958: 2228)
64 Working with Concepts

According to this model of sign, the traffic sign for “stop” would con-
sist of: the red light facing traffic (the representamen); vehicles halting (the
object) and the idea that a red light indicates that vehicles must stop (the in-
terpretant). In the work of Goguen, a complex sign (or a sign system) is a sign
that may have several levels of subsigns with an internal structure. He thus
developed Algebraic Semiotics as being a computational treatment of sign
systems. “Building on an insight from computer science, that discrete struc-
tures can be described by algebraic theories, sign systems are defined to be
algebraic theories with extra structure, and semiotic morphisms are defined
to be mappings of algebraic theories that (to some extent) preserve the extra
structure” (Goguen 1999). Describing blends as being semiotic morphisms of
sign systems (the input spaces), Goguen thus applies Algebraic Semiotics as
a way of formalizing Conceptual Blending. More specifically, he argues that
two category theory constructions, 32 pushouts and 32 colimits, give blends that
are “best possible” in a sense that involves ordering semiotic morphisms by
quality. Some examples of how this quality can be measured are:
– The most important subsigns of a sign should map to correspondingly im-
portant subsigns of its representation (more technically, this calls for pre-
serving important sorts and constructors).
– It is better to preserve form (i.e., structure) than content, if something is to
be sacrificed.
– The most important axioms about signs should also be satisfied by their
representations.
We presented Goguens’ work rather briefly and informally for much more
space would be needed and the reader would have to become acquainted with
algebraic semiotics and their application to Blending and we think that this is
an unnecessary effort with regard to understanding our work. However, when
presenting our model of blending, we will return to Goguen’s formalization
whenever there are similarities and discrepancies. There, we hope, the reader
will also gain a better insight into Goguen’s approach. As a final remark, we
have to say that although this formalization should be credited as the first at-
tempt to clarify some of CB’s aspects with as much accuracy as possible, it
leaves out some important issues, such as the optimality constraints, the ac-
tual processes of construction (composition, completion and elaboration) or
selective projection. Some of these are theoretically accounted for (or indi-
rectly implied by the formalization, such as the quality constraints example
above, which can be seen as optimality constraints), but they are not realized
in any specific way (e.g. with specific processes for generating mappings
 Mixing concepts 65

and doing selective projection), which would certainly unravel many more
problems. Finally it is noticeable that, in spite of efforts in this direction, the
formal notation for blending given by Goguen has not been applied by the
cognitive linguistics community, perhaps because it is too complex, or, more
probably, because the community still finds it ineffective in studying blends.
Veale and O’Donoghue (Veale and O’Donoghue 2000) present a compu-
tational model which relies on the metaphor interpretation system, Sapper
(presented in the next section), to establish a dynamic blend between two
domains. This blend, rather than being an independent new domain, corre-
sponds to a unifying set of correspondences of concepts from both domains,
built according to a constructor space. Therefore, although the authors argue
to the contrary, this work misses the actual creation of the fourth space, the
blend, which should have the same sort of structures as the inputs. It ends
up being the set of awakened dormant bridges that are raised during the pro-
cess (this will be thoroughly explained in section 4.2) which correspond to
a mapping rather than being an independent new domain. The assignment of
pairs of mapped elements doesn’t necessarily imply a specific blend, as some
elements may get projected, others get fused or be absent in the blend (se-
lective projection). In our opinion, Sapper generates what we later will call a
blendoid. As for emergent structure, it relies on “candidate inference” (trans-
fer of structure from source to target), which is sufficient for laying the basis
for novelty, but not for exploring it (i.e. there is no “running of the blend”).
The authors also address the optimality pressures (Veale and O’Donoghue
2000) as being by-products of the pressure to find a good isomorphism, but
we think these are a set of varied perspectives, hardly reducible to a single
measure20 . Perhaps the weakness of their approach to blending is that the au-
thors did not elaborate this idea more than in (Veale and O’Donoghue 2000),
nor give any practical results or detailed demonstrations, for then we would be
able to contrast our proposals to theirs. As we will see, Sapper, more specif-
ically its cross-space mapping mechanism, will also be useful for the work
presented here.
Lee and Barnden (Lee and Barnden 2001) focus the problem of reasoning
with counterfactuals from the point of view of their ATT-Meta system (Barn-
den 1998) (which will be briefly described in section 4.3), further analysing
it from a perspective of Conceptual Blending. A counterfactual is the rea-
soning associated to expressions of the form “If S1 then S2” (e.g. “If John
had sold his shares then he would have made a profit.”) and it implies a pri-
ori contradictions when making straight truth assertions of the constituents
66 Working with Concepts

S1 and S2 (it must be true that “John did not sell his shares”, otherwise the
“if” condition would not be necessary). Thus, to check the truth value of a
counterfactual, one has to reason by hypotheses (for which ATT-Meta is well
suited). Starting by a simple example, such as the one just given, and explain-
ing how it could be processed via ATT-Meta’s mechanisms, the authors then
proceed to more complex situations that would imply a conceptual blending,
such as “If Julius Ceasar was in charge of the Korean War then he’d have
used the atom bomb”. This would imply the blending of “Roman Empire”
and “Korean War” domain knowledge into the same space prior to analysing
the counterfactual in the same fashion as the simpler examples. The authors
do not explore in depth any of the mechanisms of Conceptual Blending de-
scribed here, as it is not the focus of their work. Conversely, we will not focus
on counterfactual reasoning throughout this book since it would imply a de-
viation from its main objectives. However, there is no reason to believe that
its results should not be extended towards this type of reasoning.

3.2.2. Some criticisms

We will now give special attention to the weaknesses and problems of the
Conceptual Blending framework. It is noticeable that the Conceptual Blend-
ing framework is still ongoing research, possibly in its early versions. Natu-
rally, it has been subject to some criticisms and its evolution to the next stages
of the cycle of research certainly depends on searching for valid answers to
these concerns .
The first and most obvious weakness of Conceptual Blending is its vague-
ness and lack of formality across its many aspects. Starting from the notion of
mental space, it is unclear what it is exactly, to which extent it is cognitively
plausible (it should be plausible, given the claims of CB as fundamental to
cognition). Indeed, the reader might have felt some discomfort with the defi-
nition we gave, and which we tried to clarify with AI examples. This problem
of definition of mental spaces becomes harder when discussing domains and
frames. In some examples, we see the blending of domains, in some others
(as happens in Seana Coulson’s book (Coulson 2000) of frames, without un-
derstanding why these are not just named as “mental spaces”. If there were a
more clear notion of mental spaces, perhaps the Optimality Principles would
become less vague (see, for example the explanation for “Intensification of
Vital Relations”21 above). This obscurity is carried over to the whole method-
 Mixing concepts 67

ology that the authors use to deconstruct blends. There seems to be no specific
set of rules for analyzing a blend other than intuition. A clear sign of this can
be observed in the generic space shown in our examples (see Appendix B),
which sometimes is a set of generalizations, sometimes it represents specific
knowledge, and other times it is just absent. Each new example may yield dif-
ferent analysis from different people, which compromises the predictability
of this framework. This, of course, shows more in complex blends, which we
tried to reduce to a minimum in the examples given (in fact, being subjective
as little as possible was one of the main restrictions for selecting examples
for experiments). In simpler blends, when it is clear that we are indeed faced
with a blend (e.g. “computer virus”, “Riddle of the Buddhist Monk”, Pegasus,
Dracula), the framework does not seem to be so controversial. Nevertheless,
it is still not clear how to distinguish a simple from a complex blend and, even
worse, it is not clear how to distinguish a blend from a non-blend. Ultimately,
it seems that everything that has a symbolic meaning is a blend (e.g. sign
language, money, machine dials, etc.), which of course leads to very extreme
claims. This reasoning falls into the same category as the claim that “every
language is metaphoric”, also a relativist perspective that, although maybe
philosophically interesting, only risks to sterilize its development unless a
valid paradigm shift is made (i.e. “since every language is metaphoric, let us
develop a different theory of language, and demonstrate its validity”).
All these problems lead to the issue of falsifiability. Since the Concep-
tual Blending framework does not predict the more complex blends and the
distinction between what is and what is not a blend is obscured, it is in prin-
ciple not falsifiable, and therefore not a theory, in a modern science sense22 .
Assuming the different perspective of research programs from Lakatos23 , the
framework of Conceptual Blending could be considered a research program,
although its belt of auxiliary hypotheses needs to be more formally defined.
In other words, these auxiliary hypotheses still need to be falsifiable.
These criticisms are intended to encourage work that, from our point of
view, is fundamental and also motivates us. We do not promise to give any
perfect blending machine or even to demonstrate that ours is the formalization
of the whole Conceptual Blending framework. To do so, it would be necessary
to solve the problems described above (what is a mental space, a frame or a
projection; what should the optimality principles and the selective projections
be about) prior to dedicating an entire thesis to it. Here, we will propose a
computational level answer to some of them.
In spite of all these criticisms, it has been claimed that Conceptual Combi-
68 Working with Concepts

nation theories (such as presented earlier in section 3.1) cannot predict more
than Conceptual Blending does, i.e., the same level of predictions with noun
noun combinations can be done with CB (see (Coulson 2000)). Indeed, as
said by James Hampton, only a small set of emergent features can be pre-
dicted by theories, which limits predictability to the more constrained and
closed world situations (i.e. in an ideal, yet unrealistic, scenario, we have two
concepts defined with a universally accepted and stable representation). As an
analytical model, it can become productive, as the examples of (Mandelblit,
1997; Sweetser and Dancygier, 1999; Coulson, 2000; Veale and O’Donoghue
2000) show how CB can contribute to Linguistics, Creative Cognition, Anal-
ogy and Metaphor.

To conclude, Conceptual Blending is as an elaboration of other works

related to creativity, namely Bisociation (in section 1.2), Metaphor (in the
following section) and Conceptual Combination24 . As such, it attracts the
attention of computational creativity modelers and, regardless of how Fau-
connier and Turner describe its processes and principles, it is unquestionable
that there is some kind of blending happening in the creative mind.

4. Metaphor and Analogy

Metaphor and analogy are two cognitive mechanisms that have been recog-
nized as underlying the reasoning across different domains25 . Because of this,
they play an indomitable role in creativity and must be discussed here. Al-
though no consensus has been reached in the current literature regarding a
clear distinction between metaphor and analogy, it is clear that their mechan-
ics share many commonalities. It is widely accepted in analogy research that
many of the problems of metaphor interpretation can be handled using estab-
lished analogical models, such as the structure mapping approach (Gentner
1983). Thus, we present a set of works that involve mapping across distinct
domains, namely SME (section 4.1) and Sapper (section 4.2). Although only
the latter has been of direct influence to our work, SME deserves particular
attention for it has been the main reference in Analogy in the past few years
and was the starting point and the benchmark for other systems, which will
also be covered in an overview (section 4.3).
 Metaphor and Analogy 69

4.1. Structure Mapping Engine

The Structure Mapping Engine (SME) of Dedre Gentner, Kenneth Forbus

and Brian Falkenhainer (Falkenhainer, Forbus and Gentner 1989) was ini-
tially built as a computational implementation of the Structure Mapping The-
ory (SMT) of Dedre Gentner (Gentner 1983)26 . In this theory (as generally
accepted in the field), analogy consists of a mapping of knowledge from one
domain (the base) into another (the target) and may be used to guide rea-
soning, to generate conjectures about an unfamiliar domain, or to generalize
several experiences into an abstract schema. Moreover, SMT is based on the
intuition that analogies are supported on relations: “No matter what kind of
knowledge (causal models, plans, stories, etc.), it is the structural properties
(i.e., the interrelationships between the facts) that determine the content of
an analogy”. Thus, analogical processing is decomposed into three stages
(Falkenhainer, Forbus and Gentner 1989):

1. Access: given a current target situation, retrieve from long-term mem-

ory another description, the base, which is analogous or similar to the
target.

2. Mapping and Inference: construct a mapping consisting of correspon-

dences between the base and target.

3. Evaluation and Use: estimate the ‘quality’ of the match. Three kinds
of criteria are involved: the structural criteria include the number of
similarities and differences; the second criteria concerns the validity of
the match; the third criteria is relevance, i.e., whether or not the analogy
is useful to the reasoner’s current purposes.

SME deals only with the Mapping and Inference stage (although also
providing a domain-independent structural evaluation). In terms of knowl-
edge representation, it differentiates between entities, predicates and dgroups.
Entities correspond to the lower level objects or constants; predicates are
higher-level primitives of three sorts (functions, attributes and relations); and
dgroups correspond to a collection of entities and predicates about them. Be-
low, we give an example of a dgroup named simple-heat-flow.
(defDescription simple-heat-flow
entities (coffee ice-cube bar heat)
expressions (((flow coffee ice-cube heat bar) :name hflow)
((temperature coffee) :name temp-coffee)
70 Working with Concepts

((temperature ice-cube) :name temp-ice-cube)

((greater temp-coffee temp-ice-cube)
:name >temperature)
(flat-top coffee)
(liquid coffee)))

To clarify, we give an example of each kind of representation: coffee is an

entity, flow is a relation, temperature is a function and liquid is an attribute. It
is clear that an attribute can be represented as a relation (e.g. (property coffee
liquid) is the same as (liquid coffee)).
SME establishes potential cross-domain linkages via match hypothesis
construction rules (MHC). These rules are programmable externally and
specify the conditions that must be met in order to create a cross-domain
linkage (known as match hypothesis(MH)).
SME constructs the cross-domain mapping by calculating the largest,
maximal collection of MH’s. “A collection is maximal if adding any addi-
tional match hypothesis would render the collection structurally inconsis-
tent” (Falkenhainer, Forbus and Gentner 1989). Being structurally consistent
means that (i) the MH’s do not assign the same base concept to multiple tar-
get concepts; (ii) if a match hypothesis MH is in the collection, then so are
MH’s which pair up all of the arguments of MH’s base and target concepts27 .
Collections are called gmaps, each containing, apart from a set of MH’s, a set
of candidate inferences (new relations that will be projected to the target, if
the gmap is chosen) and an evaluation score.
SME was able to solve many classic analogy problems, such as the Solar-
System - Rutherford Atom analogy or the Heat-Water flow analogy. In figure
4, we show its interpretation for the Heat-Water flow analogy. It can be seen
that Gmap#1 produced the correct inference that temperature causes heat
flow.
The major problem with SME is its intractability with unstructured rep-
resentations as pointed out in (Veale and Keane 1997). As the authors of
SME acknowledge, “worst-case performance occurs when the description
language is flat (i.e., no higher-order structure) and the same predicate oc-
curs many times in both the base and the target” (Falkenhainer, Forbus and
Gentner 1989). There are two major reasons for this last problem: the knowl-
edge representation is predicate centered and therefore, when predicates are
not hierarchically structured, it becomes flat, even when entities are them-
selves structured (e.g. a description of a multi-part object); SME makes an
 Metaphor and Analogy 71

Table 3. Overview of the SME Algorithm

1. Run MHC rules to construct match hypotheses.
2. Calculate the Conflicting set for each match hypothesis.
3. Calculate the EMaps (collections of entity matches) and NoGood
sets (collections of conflicting entity matches) for each match
hypothesis by upward propagation from entity mappings.
4. Merge match hypotheses into gmaps.
(a) Interconnected and consistent.
(b) Consistent members of same base structure.
(c) Any further consistent combinations.
5. Calculate the candidate inferences for each gmap.
6. Score the matches
(a) Local match scores.
(b) Global structural evaluation scores.

exhaustive search to obtain the largest mapping. In an attempt to overcome

these problems, Forbus and Oblinger proposed a sub-optimal Greedy version
of SME (Forbus and Oblinger 1990), which applies their greedy merge algo-
rithm. Roughly speaking, this algorithm iteratively cuts the search tree when-
ever it finds high quality local maxima. This quality is given by the pmap’s
“structural evaluation score”. The authors considerably optimized SME’s al-
gorithm, sacrificing its optimal results. However, again, in comparison with
Sapper (Veale and Keane 1998), the algorithm efficiency remains far from
being computationally feasible.

4.2. Conceptual Scaffolding and Sapper

Tony Veale developed a model of metaphor interpretation centered on the

contemporary theory (Lakoff and Johnson 1980), which proposes to explain
metaphors via a process of structuring one concept (the tenor) with knowl-
edge from another concept (the vehicle), with the purpose of (i) emphasizing
certain associations of the tenor over others (“my dentist is a barbarian”); (ii)
enriching the conceptual structure of the tenor by analogy with another do-
main (“the CPU is the brain of the computer”); (iii) conveying some aspect
of the tenor which defies conventional lexicalization (“the leg of the chair”,
“the neck of the bottle”) (Veale 1995). The most revolutionary assumption
72 Working with Concepts

Table 4. Complete SME interpretation of Heat-Water flow analogy(from (Falken-

hainer, Forbus and Gentner 1989))
Rule File: literal-similarity.rules Number of Match Hypotheses: 14

Match Hypotheses:
(0.6500 0.0000) (>PRESSURE >TEMP)
(0.7120 0.0000) (PRESS-BEAKER TEMP-COFFEE)
(0.7120 0.0000) (PRESS-VIAL TEMP-ICE-CUBE)
(0.9318 0.0000) (BEAKER-6 COFFEE-1)
(0.6320 0.0000) (PIPE-8 BAR-3)
oo o
oo o

GlobalMappings:
Gmap#1: (>PRESSURE >TEMPERATURE) (PRESSURE-BEAKER TEMP-COFFEE)
(PRESSURE-VIAL TEMP-ICE-CUBE) (WFLOW HFLOW)
Emaps: (beaker coffee) (vial ice-cube) (water heat) (pipe bar)
Weight: 5.99
Candidate Inferences: (CAUSE >TEMPERATURE HFLOW)

Gmap #2: (>DIAMETER >TEMPERATURE) (DIAMETER-1 TEMP-COFFEE)

(DIAMETER-2 TEMP-ICE-CUBE)
Emaps: (beaker coffee) (vial ice-cube)
Weight: 3.94
Candidate Inferences:

Gmap #3: (LIQUID-3 LIQUID-5) (FLAT-TOP-4 FLAT-TOP-6)

Emaps: (water coffee)
Weight: 2.44
Candidate Inferences:

from the contemporary metaphor theory is that metaphors belong to concep-

tual classes that are deeply entrenched in our world experience. Examples of
metaphorical concepts include ‘ARGUMENT is WAR’ (“I will defend myself
against his claims”), ‘TIME is MONEY’ (“She wasted hours in solving it”),
or ‘SAD is DOWN’ (“Don’t let yourself down”). Moreover, most non-trivial
metaphors can be reduced to complexes of simpler core metaphors, grounded
in our spatial understanding of the world (e.g. ‘ABSTRACT STATES as LO-
CATIONS’: “Bill went mad”, “Suzie went to sleep”; ‘TIME as a LOCA-
TION’: “at 5 o’clock”, “in March”).

In his model, Veale proposes two different, subsequent, steps for metaphor
interpretation: extraction of a conceptual scaffolding between the ideas
evoked by a metaphoric utterance, by identification of underlying core
 Metaphor and Analogy 73

Figure 10. Example of the scaffolding construction with four primitives (Actual and
Attempted Causality, Connect and Disconnect) (from (Veale, 1995))

metaphors; establishment of relations between those ideas for extensive ex-

planation of the metaphor, involving cross-domain transference (Sapper). The
former works top-down, while the latter, bottom-up.
The essential role of conceptual scaffolding is to build a skeleton that will
be the basis for Sapper in the search for plausible relations using domain
knowledge. Therefore, the relations built during scaffolding are not intended
to capture all the subtleties and nuances of meaning in a metaphor, rather,
they are generic guidelines for constraining the interpretation. The scaffold-
ing model takes two distinct phases: scaffolding construction followed by
scaffolding elaboration. Scaffolding construction is spatial in nature, inter-
relating the elements of the utterance through the use of primitive building
blocks (see figure 10). The composition of this structure is obtained from the
verbs involved, but other elements, such as adjectives (e.g., Big versus Small,
Lightweight versus Heavy), and explicit conceptual relations (e.g., Father,
Partner and Manufacturer) also contribute to the scaffolding.
Scaffolding elaboration labels the associations just described with partic-
ular inter-concept relation, such as Colour for connect(Porsche, Black) and
Manufacturer for connect(Macintosh, Apple-inc). These relationships are de-
rived from the interaction of the concepts governed by the scaffolding struc-
ture (see figure 11). The appropriate relationship or case is thus dependant
upon the nature of the concepts involved.
Thus, conceptual scaffolding builds a network of relations and concepts
that correspond to the skeletal interpretation of an utterance. The metaphor
interpretation is only completed after the creation of cross-domain links be-
74 Working with Concepts

Figure 11. Example of the scaffolding elaboration (from (Veale, 1995))

tween the concepts, which will allow the transfer of knowledge and, finally,
the explication of the metaphor. This is done by Sapper.
Sapper works with a semantic network that contains the information
brought from the conceptual scaffolding, also enriched with background do-
main knowledge. It applies a spreading activation based process in order to
determine novel cross-domain relations, thus reproducing much of the con-
nectionist philosophy in a symbolic framework. These are normally called
localist networks, since a distinct unit, or fixed cluster of units, is assigned to
each concept, and an activation-carrying inter-unit linkage is assigned to each
inter-concept relation (of the appropriate conductivity to capture the salience
of the relation). There are two generic aspects regarding Sapper memory that
are the most important for our purposes:
– The representation of all knowledge is equal. If not given a specific context,
all concepts and relations are equally relevant (or irrelevant), i.e. there is
no built-in hierarchy or ordering to organize the memory.
– Activation flow is entirely opportunistic. The most activated concepts will
be those that happen to be in the spots where the activation waves are
higher and in larger number, independently of what the concept actually
is or means. In other words, again, there is no a priori preference for con-
cepts or relations.
Since concepts are intrinsically dynamic, their representation should also
be dynamic and impartial. In other words, one concept can play a central
role and have a particular meaning in one context and be lateral and have a
different meaning in another context, thus it is the situation that shapes it,
not its representation. This is a well-known problem in AI, and it is clear
that Sapper does not solve it except at the very specific level of cross-space
mappings, as we will see below.
Sapper has two modes of processing, which interchange constantly as the
 Metaphor and Analogy 75

Figure 12. Example of the triangulation rule. Dashed arrow represents a dormant
bridge (from (Veale 1995))

cross-domain bridges are established:

– Structural inference is performed in a symbolic mode of processing,
– Opportunistic activation flow occurs in the connectionist mode of process-
ing, in which particular relations (the dormant bridges, as laid down in
the symbolic mode) are recognized to represent domain crossover points
between the tenor and vehicle schemata, and are thus awakened.
The structural inference is based on two rules28 : triangulation and squar-
ing. The triangulation rule states that: “whenever two concepts share an asso-
ciation with a third concept (the associations may be of different strengths),
this association provides evidence for a plausible (i.e. dormant) bridge be-
tween both schemata”(Veale 1995). In figure 12, we see a double application
of this rule: cleaver and scalpel share an association with both blood and
sharp, leading to the establishment of a dormant bridge.
The squaring rule states that, when two concepts A and B linked by a
cross-domain bridge (i.e. a dormant bridge which has already been awak-
ened) share the same association with two different concepts C and D (resp.),
then this association provides evidence for a plausible (i.e., dormant) bridge
between C and D. This rule depends on the bridges having already been awak-
ened (under the connectionist mode), thus it is considered second order. In
figure 13, a dormant bridge is created between general and brain surgeon
by the application of the squaring rule, since there is a cross-domain bridge
between command centre and brain.
76 Working with Concepts

Figure 13. Example of the squaring rule. The bidirectional arrow with ’M’ label rep-
resents a cross-domain bridge (from (Veale 1995))

In the Sapper model, activation travels in waves, each wave-form having

an amplitude (encoding the activation energy, or zorch) and a unique signa-
ture frequency. Each localist concept node is considered to possess a unique
prime resonant frequency, which is used to modulate any activation waves
that pass through this node (see fig. 14).
This propagation strategy therefore allows activation waves to be decon-
structed via prime factorization. This deconstruction reveals not only the
original source node of the wave (representing either the tenor or vehicle
of the metaphor), but also the path through conceptual space travelled by the
wave. When two activation waves cross over at the same inter-concept bridge,
Sapper is thus in a position to determine whether the waves originate at dif-
ferent source nodes in the network. If this is indeed the case, Sapper awakens
this bridge as possibly constituting a valid domain crossover point (see fig.
15). (Veale 1995; Veale and O’Donoghue 2000).

There is one final point to add to this description of the Sapper algorithm.
Inter-concept linkages also exhibit a certain resistance (the inverse of con-
ductivity) to the flow of activation energy. This provides an attenuation effect
 Metaphor and Analogy 77

Figure 14. Activation Waves in Sapper possess both amplitude (or Zorch) and fre-
quency. This signature frequency of an activation wave is the product of
the resonant frequencies of those nodes encountered by the wave in the
conceptual space (from (Veale 1995))

Figure 15. A conceptual linkage is deemed to provide a plausible match hypothesis

when it becomes a cross-over path for competing activation waves from
the tenor and vehicle concept nodes (from (Veale 1995))

in the amplitude of the waves at each node they encounter. When this ampli-
tude drops below a predetermined threshold, it ceases to propagate. Dormant
bridges have the highest (infinite) resistance, until they are awakened and at-
tributed a resistance consistent with the structural evidence brought by the
wave.
Finally, we give an example of the interpretation of “My surgeon is a
butcher” generated by Sapper (figure 16) and the returned output for “The
General is a Surgeon” (in table 5, figures on the left represent the conductiv-
ity of the cross-domain linkages).
To conclude, while Conceptual Scaffolding is knowledge dependent and
needs detailed specification and coding of the core metaphors it analyzes,
Sapper is an algorithm that finds a 1-to-1 mapping between two domains
which, although not guaranteeing the optimal solution (whatever the criteria
chosen), is computationally tractable and does not demand big compromises
78 Working with Concepts

Sacred Edible
Flesh

isa
at
tr isa tr
at

Human Flesh M Meat

Highly Paid Lowly Paid

affe
c t
affe

ct
attr Surgeon Butcher attr
M Clumsy
Skilful att
r
attr a t tr
Blood attr
r

att
att

r
Precise Careless
Location

isa perform
Bright isa Dark
attr M attr
att r
Operating r att Abattoir
attr

attr
Theatre
Hygiene
Safety Danger
Activity
Slaughter
Surgery isa isa
attr

attr
M
r Blood attr
attr
att attr
Surgical Pain
tr
at
at

att
tr

r
r att
Death
part

part

Scream
Clinical

Scalpel Cleaver
Incision Chopping
M perf
rm
perfo attr Delicate Hefty
att
r orm
*
att r
tr

at

r att
at

tr
attr

attr

Sharp

Sterile
Dangerous
M

Precise Clumsy
isa
isa

Accurate

Figure 16. Interpretation for ’SURGEONS are BUTCHERS’ (from (Veale 1995))
 Metaphor and Analogy 79

Table 5. Output from Sapper for ’GENERALS are SURGEONS’

0.98 If General is like Surgeon
0.76 Then 18th-Century-General is like Saw-Bones-Doctor
0.95 and Soldier is like Patient
0.81 and Casualty is like Corpse
0.91 and Bombing-Raid is like Surgery
0.93 and Atomic-Bomb is like Radiation-Therapy
0.75 and Nerve-Gas is like Disinfectant
0.88 and Enemy-Army is like Cancer
0.78 and Enemy-Soldier is like Cancer-Cell
0.5 and On-Target is like Precise
0.7 and Snub-Fighter is like Scalpel
0.5 and Military-School is like Medical-School
0.17 and Battlefield is like Operating-Theatre
0.17 and Military-Uniform is like White-Smock

regarding knowledge representation or configuration.

Veale and Keane (Veale and Keane 1997) presented a very interesting
comparison of Sapper with SME (Greedy and non-greedy versions) and
ACME (Holyoak and Thagard 1989) (described later). This analysis vividly
reports how well Sapper behaves in comparison with its peers, namely con-
cerning tractability issues. Apart from this paper, the interested reader should
also inspect Ferguson et al’s reply (Ferguson, Forbus and Gentner 1997).

4.3. Others

As with other surveys (e.g. (French 2002)), we classify systems on metaphor

and analogy into three types: symbolic (ATT-Meta, SME and MIDAS), con-
nectionist (LISA) and hybrid (Drama, Copycat and Sapper). Once again, this
overview undoubtedly leaves out many systems. Our purpose is to provide a
broad picture of the field, and eventually lead to a synthesis of the features
and problems that characterize it.
ATT-Meta (Barnden 1998; 1999) is a rule-based system for metaphorical
reasoning initially projected to process mental states in discourse. It there-
fore focuses on specific types of metaphor schemata, such as ‘MIND PARTS
as PERSONS’, and applies built-in commonsense models in order to inter-
80 Working with Concepts

pret sentences like “One part of John was insisting that Sally was right”.
ATT-Meta then triggers rules that propose possible interpretations for the
metaphor, according to different pretences29 . It isolates pretences (thus avoid-
ing logical inconsistencies) within cocoons, enabling the simultaneous con-
sideration of several different, possibly conflicting, hypotheses. This system
allows representation of uncertainty in its knowledge, which will then serve to
evaluate the truth probability of a pretence (and propagating this probability
to other, dependent, pretence cocoons) and propose a plausible interpretation.
ATT-Meta does not itself deal with natural language input directly. Therefore,
a user supplies hand coded logic formulae that are intended to express the
literal meaning of small discourse chunks (two or three sentences)(Barnden
1998).
Another work on metaphor reasoning is MIDAS (Martin 1990). As with
Conceptual Scaffolding, MIDAS approaches interpretation with the assump-
tion that there is a set of core metaphors. These are stored in a knowledge
base that is continually augmented with extended metaphors, which derive
from the core metaphors. Each metaphor is represented by a structure called
metaphor-sense, which contains a source, a target, and a set of associations.
An association is represented by a metaphor-map, which links a source con-
cept to a target concept. MIDAS interprets a metaphoric utterance by retriev-
ing the most similar metaphor and adapting it to the current situation. In this
sense, it works as a Case-Based Reasoning system, whose learning ability
relies on the storing of newly adapted cases.
John Hummel and Keith Holyoak proposed an artificial neural-network
model of relational reasoning, LISA (Learning and Inference with Schemas
and Analogies) (Hummel and Holyoak 1997), which uses synchrony of fir-
ing to bind distributed representations of relational roles (e.g., the roles of
opposite-of(X, Y)) to distributed representations of their fillers (e.g., black
and white). Thus, a proposition corresponds to a pattern of activation. LISA
has a Working Memory (WM) containing the target (and the source, after re-
trieval) being investigated; and a Long Term Memory (LTM), which holds the
candidate sources. When a target is specified in WM, its pattern of activation
triggers the retrieval of the appropriate source proposition from LTM, which
is the one that is better synchronized with the pattern of activation of that tar-
get. These two memories have distinct representations. WM comprises a dis-
tributed representation (as is traditional on pure connectionist system), while
the LTM is localist (as with Sapper). For example, if the WM contains the
target proposition “Beth sells her futon to Peter”, then it may retrieve an anal-
 Metaphor and Analogy 81

ogous source proposition (e.g. “Bill sells his car to Mary (and so Mary owns
the car)”). When two analogous propositions are met in LISA’s WM, their
co-mapped constituents are co-activated in synchrony (Bill to Beth, car to fu-
ton, etc.) and it is possible to transfer inference from the source to the target
(i.e. “Peter owns the futon”), following the same activation procedure (“Pe-
ter” is co-activated with “Mary”, who “owns” a “car”, which is co-activated
with “futon”). LISA’s main limitations concern the WM memory sizes and
LTM representation issues. Indeed, the WM can only store one proposition
at a time, which forbids solving complex analogies. Moreover, the built-in
representation of LTM makes LISA an uncreative system with low flexibility,
since it demands the explicit coding of each proposition.
Drama is a system that aims to integrate semantic and structural informa-
tion in analogy making (Eliasmith and Thagard 2001). It has a set of particu-
larities that make it unique among its peers. The foremost is its application of
holographic reduced representations (HRRs) (Plate 1994) memories, which
allow the distributed, vector-based, representation of concepts and relations
in Drama. The storage operation of a vector in a HRR is called convolution,
while the retrieval operation is called correlation. HRRs allow the convolu-
tion of large amounts of information in the same memory space, but the more
they store, the lower reliability they will provide in correlation. It is then
necessary an error-cleaning mechanism. It is claimed that HRRs are cog-
nitively plausible models of memory (Eliasmith and Thagard 2001). Other
systems also apply distributed representations such as neural networks (e.g.
LISA), so a thorough comparison should be made to understand which one
is better in analogy contexts. In Drama, each ground concept is attributed a
random vector that is then stored in a HRR, along with its semantic infor-
mation (properties and ISA relations, each defined as an independent vec-
tor). Domain structure (relations between different concepts) is also stored in
the HRR in the form of vectors. Given the source and a target proposition
vectors, Drama starts the analogy-mapping by obtaining their similarity (via
vector dot product). When they are sufficiently similar (higher than a thresh-
old), it then proceeds to the constituents. For each pair of similar constituents,
Drama builds a node in a network (the mapping network), and establishes
links between nodes that participate in the same relation. This latter process
is the same as ACME’s (Holyoak and Thagard 1989) algorithm for analogy
mapping: in ACME, the algorithm starts by establishing a network of map-
ping pairs, each node containing a pair, each pair linked to other pairs. Using
the LISA example above, ACME (or Drama) would initially build nodes for
82 Working with Concepts

“sells” with “sells” then for “Beth” with “Bill”, “car” with “futon”, etc. each
of these nodes being co-activated. It could also generate competing mapping
nodes in the network (e.g. “Beth” with “Mary” - both are women) which
would have little co-activation with the former nodes. Then, with a spread-
ing activation process (as in Sapper), it would select the mapping sets that
best satisfy the constraints of similarity, structure and purpose, as defined in
(Eliasmith and Thagard 2001) and (Holyoak and Thagard 1989). In theory,
Drama can integrate both structure and meaning, which would be a major
breakthrough in analogy research, but, since the ground concepts are given
random vectors, the meaning is entirely dependent on the property and ISA
relations, which end up as being structural knowledge as any other relation.
Indeed, although the authors treat these relations differently, they do not cor-
respond to the notion of meaning that they advocate. More specifically, the
problem lies in the randomness of encoding the ground concepts and their
resulting similarity (e.g. in one run, “dog” and “cat” can be more similar than
“dog” and “freedom”; while in the following one, the opposite can happen
without any particular reason). The authors claim that this is coherent with
the psychological differences between people, but randomness does not seem
to be a good model for it. A proper solution would be to learn the meanings,
as also suggested in (Eliasmith and Thagard 2001). However, this learning
algorithm is by itself a challenge.
Also unique in many aspects, Copycat (Hofstadter and Mitchell 1988) is
a system for solving puzzle analogies (such as “abc→abd::xyz→?”) as al-
ready presented in section 2.2.2. This system has many nuances and has been
deemed an example of computational creativity, as well as its related fam-
ily: Tabletop, Letter-Spirit and Metacat (Hofstadter 1995). Nevertheless, it
has been criticized as only being able to work on a very specific, exhaustively
defined domain. In fact, while an omnicompetent Slipnet is theoretically plau-
sible, in practice, serious resources are necessary even for simple domains.
The issue of knowledge representation has been evoked constantly as a
central problem in any of these systems. Some approach it by focusing on
specific domains (ATT-Meta, Copycat), some try to cover generic knowl-
edge (Sapper, SME, MIDAS, LISA, Drama). Some rely on structure mapping
(SME, Sapper), some on axiomatic inference (ATT-Meta, MIDAS), some try
integrating both (Drama, Copycat). Nevertheless, it is clear that each one is
ultimately dependent on built-in domain representations (the exception being
Drama, which solved the problem with randomness). The path now taken in
the area seems to lead to hybrid approaches, both in terms of paradigm (sym-
 Metaphor and Analogy 83

bolic and connectionist) and in terms of inference mechanism (axiomatic and

structure mapping).
The majority of the works (with the exception of Copycat) allow only 1-to-
1 structure alignment between domains, but it has been pointed out that many-
to-one mappings may be useful, both in metaphor and in analogy (Falken-
hainer, Forbus and Gentner 1989). Even more, structural alignment easily
falls prey of the representation of domains. For example, having a source
with “isa(dog, pet)” and “isa(pet, animal)”, and a target with “isa(cat, ani-
mal)” and “isa(animal, entity)”, it would not yield the analogy of “dog” with
“cat”, rather “cat” would be mapped to “pet”, and a possible analogical infer-
ence could be “isa(dog, cat)”. This once again raises the problem of represen-
tation. As we will see in chapters 5 and 6, the same questions can be raised
for this work.
To conclude, there is general agreement that metaphor and analogy are
cognitive mechanisms that can uncover aspects unforeseen, by bringing
knowledge from one source to a target and thus making predictions, solv-
ing a problem or even by expressing concepts that have yet no conventional
meaning. So far, the computational approaches to these mechanisms have es-
sentially relied on cross-domain mappings and structure alignment, which is
not to say that other approaches aren’t worth of attention.
Chapter 4
A Model of Concept Invention

In this chapter, we will take a top-down approach, presenting and discussing

the requirements for an abstract creativity model, and then progressing to-
wards a formal model of concept invention that will be the subject of an
implementation presented in the next chapter. The reader should understand
that, on the way to, and particularly when arriving at the actual implementa-
tions, many choices have to be made, both in terms of what aspects to focus
on (e.g. bisociation vs re-representation) and in terms of practical decisions
(e.g. implementing algorithms, choosing representations). We will try to jus-
tify each decision whenever any of these choice points arise.

1. A Creative General Problem Solver

In chapter 2, questions regarding creativity were raised. By doing so, we in-

tended to provide the reader with the set of principles followed in the con-
struction of the model for computational creativity proposed here:
– Knowledge. It has been emphasized that there is rarely creativity without
knowledge and that both quantity and quality should be treated as equally
important. A model for creativity should consider a heterogenous knowl-
edge base, in the sense that it should not include solely the typical knowl-
edge for solving a specific problem, but instead many different domains
and perspectives towards more than one problem (Weisberg 1999).
– Re-representation. It is also important to be able to understand the existing
range of knowledge according to different points of view. A model for
creativity should be able to change the representation of a concept without
losing its meaning (Karmiloff-Smith 1993).
– Bisociation. The notion of bisociation is connected with cross-domain
transfer, to the ability to find unprecedented associations. A model for cre-
ativity should be able to find and explore associations between distinct
knowledge structures, namely structures that seem apparently distant and
unrelated (Koestler 1964).
– Meta-level reasoning. The ability to reason about reasoning is also a trace
of creativity. This aspect is perhaps the most difficult to specify, but should
86 A Model of Concept Invention

also be taken into account. As well as being able to process the knowl-
edge, a model of creativity should be able to process its own processes
of processing knowledge, preferably without having to employ different
techniques for each level of abstraction (Colton 2001; Wiggins 2001).
– Evaluation. An indisputable part of the creative process has to do with eval-
uation, both in terms of the self and of the society. A model for creativity
should be able to do self-assessment and react to external evaluation (Csik-
szentmihalyi 1996; Boden 1990).
– Interaction with the environment. No model of creativity should be de-
signed without taking into account the environment. Indeed, some re-
searchers have emphasized that creativity can only be perceived against
a context, which includes the individual (as a producer and recipient), the
society, the History, the motivations, in other words a set of aspects that lie
outside the scope of the new concept or idea being considered (Csikszent-
mihalyi 1996).
– Purpose. There is always a purpose in any creation, even though it may
be sometimes extremely subtle. We do not agree with the argument that a
creative system does not have to be goal-oriented. Creativity happens as a
necessity rather than as a purposeless activity, whether for satisfying some
fuzzy aesthetic preferences or for solving a practical problem (Amabile
1983).
– Divergence/convergence. One of the main conclusions taken from chapter
2 is the existence of two modes of thinking, the divergent and the con-
vergent, both important to creativity in different aspects. Thus, a model
of creativity should consider both divergent thinking, which is when free-
association is sought, a less controlled search is allowed, constraints can
be broken and inconsistencies may be generated; and convergent thinking,
which is methodic and driven by rationality (Guilford 1967).
– Ordinary processes. A final aspect to raise is that there is no reason to
believe that, underneath a creativity model, there need be processes that are
special or fundamentally different from the ones applied in non-creative
reasoning. Furthermore, there is no reason to argue that bisociation and
divergent thinking are not grounded in the same cognitive processes as any
other cognitive phenomenon or mode of thought. In other words, all these
are manifestations of intelligence, with divergent thinking and bisociation
being the ones that are more commonly identified with creativity (Guilford
1967; Csikszentmihalyi 1996; Koestler 1964; Boden 1990; Finke, Ward
and Smith 1992).
 A Creative General Problem Solver 87

Environm ent ( ex te r n a l e va l u a tio n , in te r a c ti o n )

Multi-domain Knowledge Base

Re-representation
Bisociative Reasoning
mechanism Mechanism

Pragmatics module (goal statement, self evaluation, preferences)

Figure 17. The Creative General Problem Solver

From these eight principles, we propose a Creative General Problem

Solver (after an analogy to the General Problem Solver of Ernst and Newell
(Ernst and Newell 1969)). In doing this purely philosophical exercise, we
have two intentions: to provide a model that summarizes all aspects and to
focus on the relationship between Creativity and AI.
In figure 17, we show a model that considers the many aspects referred
to above. The reasoning mechanism, perhaps the least obvious in terms of
its internal workings, should be responsible for controlling the whole system
and for doing the search according to the goal, preferences and evaluation
given by the pragmatics and environment modules. The working mode of the
system (either divergent or convergent) should depend on the use that the rea-
soning mechanism makes of the bisociative mechanism and re-representation
modules. A purely convergent mode would not use any of these modules
while a purely divergent mode would apply them for every step in the search.
The re-representation module provides different alternatives of representation
of the knowledge in the multi-domain knowledge base, while the bisociative
mechanism is expected to find and propose associations between any two dis-
tinct structures in that knowledge base. A particular aspect of the knowledge
base (not explicit in the diagram) is that it should also contain the specifica-
tion of every other module; in other words, the whole model is itself part of
the knowledge base. This is definitely the most complex aspect and relates to
meta-level reasoning. Indeed, if the system is to reason about its reasoning, its
88 A Model of Concept Invention

knowledge base must also consider, directly or indirectly, the representation

of all its processes. The question is how this can be implemented and where
this recursion ends.
Now, analyzing this model from an AI perspective, the basic question
arises: Why do we call it creative? Is it not yet another AI model? Is it not
just search over a complex space, which contains the solutions given by the
plain knowledge base, added with its several different re-representations, the
associations that result from the bisociation method, and the description of
the processes themselves? From the (behaviorist) perspective of the prod-
uct, one can call it creative when, from this search over a complex space,
novel and useful ideas tend to emerge. From the (cognitive) perspective of
the process, we can argue that, if the set of principles listed above are in-
deed connected to creativity, then such a model should produce more novel
and useful solutions to a problem than if the same problem was given to a
classical AI model (which would essentially have no bisociative mechanism,
meta-level reasoning or re-representation module, and would have a much
more focussed knowledge base). These proofs seem impossible to demon-
strate formally, so we are left to explain exactly what such a model brings
new to the area. Its possible contribution to AI is that it explicitly incorpo-
rates for the first time all those aspects mentioned, and, most of all, the asser-
tion that creativity should be considered as one side of problem solving and,
therefore, of intelligence. To clarify some more, such a model could solve a
problem in a convergent manner, i.e. recurring to the knowledge specifically
directed towards the problem (let us call it problem-specific knowledge), in-
dependently of how many times the same problem and solution had been
brought up. It could also try to find a different solution, also with problem-
specific knowledge, possibly seeming more creative. Solving in a divergent
manner would imply the search for other associations and/or representing
knowledge differently or even changing its own processes. In this case, when
finding previously unseen solutions, the result may have a higher probability
of being deemed creative by an observer. In any of these cases, we deal with
a classical view of intelligence as problem solving, in some cases resulting in
creative solutions.
The model just presented is very abstract and consists essentially of a set
of guidelines to be explored further either by us (some are covered in more
detail later in this book) or by others in future work.
In the next section, we descend one level more in our top-down approach
to develop a creativity model. More specifically, we focus on a set of features
 Description of the model 89

already considered, namely bisociation, an heterogenous knowledge base,

meta-level reasoning, convergent/divergent modes of thinking and purpose.
These are the foundations for our model about concept invention.

2. Description of the model

We now present our model of concept invention. It is focussed essentially

on bisociation, an heterogenous knowledge base, meta-level reasoning, con-
vergence/divergence and purpose. Therefore, we pay little attention to the as-
pects of interaction and evaluation and leave out re-representation. Interaction
will be reduced to goal statement and configuration, and evaluation will be
based on self-assessment. These choices result from a priority given to sub-
ject versus society, and therefore focussing on its inner processes. Because
the issue of re-representation deserves an entire thesis to itself, we decided
not to consider it further. Nevertheless, we will return to it when we think it
relevant in this text.
Before giving details, we would like the reader to imagine an ideal sce-
nario where a system has been given a goal to reach. This goal could be
something like “the specification of a flying transportation object”. The sys-
tem has not enough knowledge of airplanes, or physics, or something that
could lead by sound processes to reach directly this goal. Or it has indeed
the necessary knowledge, but the complexity of the search space is too big
to reach the goal in a reasonable amount of time. It could then enter a mode
of divergence, in which combinations between concepts in memory would
be made, always checking if something similar to the goal is achieved. After
reaching the most promising idea (say, after spending a while in divergence,
it had found “a bird with a box connected”), and if still not achieving the goal
(e.g. “the box is too heavy”), this system would then return to a convergent
mode, in order to elaborate the idea to reach a satisfactory solution. If, in the
end, a good solution had still not been found, the system could also try to
invent new ways of combining concepts, of elaborating, and of searching i.e.
it would try to improve its own processes. Our model of concept invention is
concerned with the divergent part of this scenario. In figure 18, we show its
diagram. It has six modules:
– Multi-domain knowledge base. The knowledge base follows the exact
same principles as described for the Creative General Problem Solver.
90 A Model of Concept Invention

U se f u ln e ss
Multi-domain Knowledge Base

Evaluation
Bisociative mechanism Reasoning Mechanism

Mapping function Divergent Strategy

No v e lty
Transfer operation Convergent Strategy

Elaboration Goal

Figure 18. A Model of Concept Invention

– Bisociative mechanism. The bisociative mechanism starts by finding map-

pings between concepts and creates a new concept. Then, from these map-
pings, it transfers knowledge from each of the co-mapped concepts to the
new, bisociative, concept. The mappings do not have to rely on similarity:
they can represent conflicts that are striking, surprising or even incongru-
ous (as suggested by Koestler (Koestler 1964).
– Reasoning mechanism. The reasoning mechanism applies two strategies.
The divergent strategy makes use of the bisociative mechanism to generate
new concepts and picks the ones that get a better evaluation. The conver-
gent strategy makes use of the elaboration mechanism to generate better
concepts from the ones resulting from the divergent strategy
– Evaluation. The Evaluation module returns the measure in which, accord-
ing to a goal, a given concept satisfies the criteria of novelty and usefulness.
– Elaboration. Elaborating or adapting means reworking a concept to comply
with context and domain-dependent constraints. In other words, the Elabo-
ration module is concerned with eliminating inconsistencies and complet-
ing a concept with valid knowledge.
– Goal. The Goal should be given externally, it defines the purpose of the
concept being sought.
This model was first sketched and presented, with slight but irrelevant
differences, in two papers: “Modelling Divergent Production: a multi domain
approach”, presented at ECAI’98 (Pereira, 1998), and “Wondering in a Multi-
 Description of the model 91

Domain Environment with Dr. Divago”, presented at CSNLP’99 (Pereira and

Cardoso 1999). At neither time was it formalized in detail. This will be done
in the following pages.
For those readers unaware or not interested in this formal specification, we
will also describe each idea informally. In order to perceive the underlying
ideas, it should not be necessary to understand every technical detail.
In this formalization, we will borrow some definitions from Geraint Wig-
gins (Wiggins 2001) (see section 2.1.1), namely the universe, U , of concepts,
the language L and the traversal strategy, T .
Thus, let the alphabet A contain all possible atomic symbols (constants
and variables) conceivable. Let us also define a process, p, by which we can
compose elements of A in order to get higher order knowledge structures,
the language L , which may comprise predicates and functions (a predicate
P would take the form P(x1 , x2 , ...xn ), with P, x1 , x2 , ...xn being symbols of the
alphabet A ). To simplify, we assume that these higher order knowledge struc-
tures have the form of facts and rules (as in a logic program, e.g. (Leite 2003)
and that the process p is a generative grammar that allows the generation
of all possible concepts (i.e. logic programs) of language L . Therefore, this
will be our universe U , of concepts, which will contain, as well as other con-
cepts, all the rule sets T , E and R, and their associated sets,  R, T , E ,
[[R]]and [[E ]]. Remember that, according to section 1, each concept is defined
as a micro-theory (formally, a logic program). Informally, let us imagine the
universe U of all possible concepts representable by a language L . Follow-
ing the principles in section 1, our model must consider a knowledge-base
with many different domains and, in order to allow meta-level reasoning, the
description of its own controlling processes. To sum up, the knowledge base
in our Model of Concept Invention should contain concepts from the domain
for which we intend to invent concepts, from other domains, and the descrip-
tion of the model itself. The logic program (micro-theory) that describes a
concept specifies its relationship with other concepts as well as its inner char-
acteristics. The set of concepts contained (intensionally or extensionally) in
a knowledge base KB is defined as UKB , with UKB ⊆ U . A knowledge base
corresponds to a set of concepts that cohabit the same physical or virtual
memory space. It should be a model of an individual’s own complete knowl-
edge. A knowledge base is multi-domain if it contains concepts from more
than one domain.
Let a domain D correspond to a set of concepts from KB (so D ⊆ UKB )
such that all of them relate to a unique, underlying concept. Of course, this
92 A Model of Concept Invention

invites the extremes that the KB itself is a domain (“everything is member

of KB”); and that each concept is a domain (“Each piece of knowledge that
describes the concept is related to that concept”). The notion we intend to
bring is that a domain incorporates knowledge relating to the same subject.
For example, a set of musical pieces, along with style rules corresponds to
a domain of music, while a set of recipes with data describing available in-
gredients is a domain of cookery. As with our own knowledge, domains can
encode different levels of expertise, detail, and so on. This notion of domain is
rather imprecise and it is not conclusive for the model we are presenting. We
introduced it mainly to assert that the knowledge base comprises potentially
disparate kinds of data that are commonly associated to knowledge about dif-
ferent subjects.
According to Koestler, Guilford, Fauconnier and Turner, and many other
theorists already referred to in this document, the associativity between con-
cepts is fundamental in creativity. This takes us to the next definition, the
mapping function:
φ : U × U −→ {0, 1}
This definition implies that any element from U can be mapped to an unde-
termined number of other elements (more specifically, whenver φ (x, y) is 1,
this implies that x can be mapped to y). φ should obey the following axioms:
1. φ (x, y) = φ (y, x) : x, y ∈ U , i.e. φ is symmetric.
2. φ (x, Ø) = 0 : x, Ø ∈ U , where Ø is the empty concept.
The empty concept has no content (e.g. a logic program with no facts or
rules) and can be represented by a constant (in Divago, this will be the atom
nill).
We will call the function that maps elements from two domains a cross-
domain mapping function:
φD1 ,D2 : D1 × D2 −→ {0, 1}, where D1 and D2 are distinct domains.
The act of bisociation is not completed until a novel concept emerges. In
the many examples given (of bisociation, blending, conceptual combination),
knowledge is transferred from each co-mapped element to the novel concept.
This is where, in Koestler’s model, the act of creation happens, and where,
in Conceptual Blending and Conceptual Combination, emergence starts to
happen:
ω : U × U −→ U
 Description of the model 93

ω , the transfer operation, can be defined as:


Ø : φ (x, y) = 0,
ω (x, y) =
k otherwise, with some k ∈ U .

Informally, the transfer operation is an operation over two concepts (x and

y) such that, if there is a mapping between them, a concept k (from U ) is
created. There is no way to specify this operation in more detail, since there
are many different accounts for how concepts are combined together. In the
next chapters, we will propose a possible version whose results will also be
demonstrated. ω is not necessarily deterministic, i.e. k may correspond to a
different concept each time ω is applied. Two other axioms for ω also follow:

1. ω is not symmetric. I.e., ω (x, y) = ω (y, x) is not necessarily true.

2. ω (x, Ø) = Ø

The set Ω contains all possible bisociations within U . We call it the biso-
ciation set:
Ω = {k : k = ω (x, y), x, y ∈ U }

When ω is applied to two entire domains, we obtain the set ΩD1 ,D2 . We call
it a domain bisociation:

ΩD1 ,D2 = {k : k = ω (x, y), x ∈ D1 , y ∈ D2 }

The bisociation set thus contains all possible bisociations for a knowledge
base KB. In our model, this set contains the structures that result from what
Guilford called divergent production, also to which (Finke, Ward and Smith
1992) called the pre-inventive structures, in the Geneplore model.
The choice or ordering of the bisociation set can only be made if there is
a goal to be reached. The agent should be looking for something, otherwise
there would be no particular reason for picking one element from the bisoci-
ation set, i.e. to explore pre-inventive structures or to converge on something
interesting. Thus we define the set Ug of all possible expressed goals, based
on Lg , such that Lg ⊆ L . A goal can range from very specific requirements
for a problem (e.g. a set of design requirements) to abstract (e.g. achieve
balance in a picture) and vague requirements (e.g. need for joy). Associated
evaluation functions, which test whether a concept does or does not fulfil the
94 A Model of Concept Invention

criteria, and verify its novelty, must underpin any of these goals. Thus, we
have two functions, novelty and usefulness:
nov : U × Ug −→ [0, 1]
The function nov returns the novelty of a concept w.r.t. a goal g. In the typ
function of (Ritchie 2001), which should be the inverse of novelty, the goal
was implicitly considered, but we think goals should not be singular or en-
coded in functions, they should also be seen as concepts and be members of
U . This means, in the notion of concept followed here, that goals should also
be expressed as micro-theories. Another implication is that goals themselves
can be bisociated. The usefulness is given by the other function:
use : U × Ug −→ [0, 1]
Again, the accomplishment of a goal is as fundamental as the concept itself
in order to assess its usefulness. Something is only useful (or appropriate) if
it is seen in context.
Now examining the bisociation set, its size may vastly increase and the
novelty and usefulness of its members may vary extensively. In other words,
the search space for getting good concepts can become extremely complex.
Following (Wiggins 2001), we also propose the traversal set, T , which em-
beds the strategy used by an agent to traverse a search space. We assume that
T ⊆ U , i.e. the strategy is also defined as a set of concepts from the universe
U , also implying that the same operations could be applied to the strategy it-
self. We propose no practical realization for this meta-level reasoning in this
work, but we want to stress that it is a fundamental aspect if we want to reach
the limits of computational creativity.
The bisociation set is traversed by a strategy Td . In the traversal of
these pre-inventive structures, definitive values for novelty and usefulness are
sometimes hard to assess, so priority to other measures of interest (not nec-
essarily driven by the goal) may also be applied, such as diversity, simplicity,
re-representation potential, etc. In other words, although we are only consid-
ering the functions of novelty and usefulness here, other factors may be of
importance in the selection of elements from the bisociation set. Td should
thus be understood as the divergence strategy, which should cognitively cor-
respond to the act of wandering for possible solutions, or inspirations, to a
problem.
Also following Wiggins approach, we suggest the function operator, 
., ., . , which selects elements of U from existing ones. Wiggins proposed
 Description of the model 95

an application of this operator as  R, T , E , under the assumption that

a search strategy T would traverse a space (partially) ordered by the rule
sets R and E . In our case, we propose the traversal of the bisociation set,
Ω, partially ordered by the functions of novelty and usefulness. The implied
correspondences (use ↔ R and nov ↔ E ) are no coincidence, rather we can
assume they are instantiations: from [[R]], we select those that are useful for a
goal; from [[E ]], we prefer those that are novel30 . In so doing we avoid some
of the vagueness implied by the definitions of R and E and redefine the step
of the space traversal to be:

xi+1 = use, Td , nov  [xi ]

Informally, what we mean exactly here is that we expect our divergent

search procedure to consider both novelty and use f ulness in the traversal of
the concept spaces. Moreover, this traversal is cumulative in the sense that it
counts with the concepts already inspected (in the same way that a composer
will not invent the same exact idea twice, since in the second time he will
have the previous concepts in memory). The application of this operation
repeatedly to achieve all possible concepts would lead us to the generation
of the complete divergent strategy set. Formally (again following Wiggins
notation), this set corresponds to:

 use, Td , nov  [Ω]

Notice that we apply the function directly to the set Ω, instead of giving the
“starting symbol”, the empty concept ⊥. The intent is to explicitly assert that
this search is made within the set Ω, instead of the whole universe U . Also
notice that the resulting set will be a subset of the set Ω itself. Informally,
after defining the set of all possible bisociations (certainly a very large set),
one needs to choose good ones. This can only be done via criteria such as
novelty and usefulness in conjunction with a search strategy.
We have reached the point where a (set of) concept(s) is found, still in
its “pre-inventive” state. In other words, it would be expected that further
exploration is needed in order to arrive at a proper answer to the goal(s). This
corresponds to the convergent phase, also previously referred. Let us now
have the function θ , called elaboration function:

θ : U × U −→ U
96 A Model of Concept Invention

This first argument of this function is the concept to be elaborated, while

the second corresponds to the method used. Below, there are three kinds of
elaboration:
– θR : C × R −→ C , where R ∈ U is a rule. This is called rule-based elabo-
ration, which happens when a rule or a set of rules (external to the concept)
is applied for elaboration. These rules can be heuristics, causal rules, what-
ever kind of production rules available.
– θC : C × C −→ C , where C is a concept. This internal logic elaboration
consists of applying reasoning methods exclusively taking into account the
concept’s micro-theory. Examples of these reasoning methods are deduc-
tion, induction and abduction within the micro-theory of C.
– θC C : C ×C1 −→ C ,C = C1 such that C,C1 are concepts. This is the cross-
concept based elaboration, where the first concept is elaborated by asso-
ciation/comparison to other concepts in the knowledge base. Sometimes,
new knowledge or structure is added or removed to a concept by compar-
ison to other concepts. A special case of this is the cross-domain based
elaboration, when C is elaborated by concept(s) from a domain that differs
from that in which C is integrated. This kind of elaboration is often used in
analogy, when knowledge from one domain gets projected onto the other.
We also add an axiom for θ :
1. θ is not symmetric. I.e. θ (D, K) = θ (K, D) does not necessarily hold.
We can now define the set Θ of elaborated concepts:
Θ = {θ (C, K ) : C ∈ use, Td , nov  [Ω] ∧ K ∈ U }
Informally, the set Θ is constructed by applying all the possible elabora-
tion operations to all concepts generated before by the divergent strategy. As
with the bisociation set, the set of elaborated concepts can be ordered by the
functions of novelty and usefulness, yielding the search space that is to be
traversed by Tc ⊆ U , the convergent strategy. The final result, the ordered
set of concept inventions, is given by:
 use, Tc , nov  [Θ]

3. Discussion

This model implies the interaction of three concept sets, all belonging to the
universe U : the set UKB , consisting of all the concepts implicitly or explicitly
 Discussion 97

defined in the available knowledge base; the set Ω, which contains all the
bisociations generated; the set Θ, comprising all the possible elaborations
of the elements from Ω. Two more sets should be considered, regarding the
definition of the search strategies, UT and the goals, Ug . In figure 19, we
depict all these sets as well as the directions taken by the search strategies. We
should remember that, since this is a purely theoretical analysis, one should
not read too much into the exact position and size of the sets. Our intention
here is to relate our model to each of the sets and their intersections.

Figure 19. An analysis of the model presented

We defined two sorts of strategies, the “simple” strategies (Tc and Td )

and the “ideal” strategies (Tc and Td ). They should represent a continuum
of possibilities. The “simple” (or rather computationally “realistic”) strate-
gies consider a clearly bounded concept space. The path depicted show what
would be expected of a plain application of the model just described. It starts
from the universe of known concepts, UKB , then applies bisociation, thus in-
specting the set Ω with the divergent strategy. The convergent strategy will
then take the search within the set Θ. The “ideal” strategies should be able to
98 A Model of Concept Invention

diverge more, considering the sets UT and Ug , as well as having the freedom
to jump off the bisociation space as it was defined (e.g. due to a change in its
own goal and strategy). Again, this is more a dissertation than a practical pro-
posal. Indeed, this leads us to consider the “simple” strategies as a realizable
step towards that ideal. The intuition behind the two sides of this continuum
is the range between day-by-day creativity and the revolutionary creativity, or
the big “C”. We suggest that divergence as well as convergence are constantly
present in daily problem solving, although only in special situations does it
become necessary to diverge considerably, i.e. to apply a Td kind of strategy.
A somewhat different analysis can be made regarding the expected weight
of novelty and usefulness in the application of the strategies (figure 20). Here,
we have a closer match with the diagram proposed for Wiggins’ formaliza-
tion, in section 2.

Figure 20. An analysis of the model presented by applying the framework of Wig-
gins, modified by instantiating R with use and E with nov

We have placed the universe of concepts defined by the knowledge base,

UKB , in the center of  use, T , nov . In fact, as well as being totally avail-
 Discussion 99

able for the search method, UKB should contain a variety of concepts, ranging
from the useful to the non-useful, from the novel to the typical31 . We also as-
sume that Tc would be directed towards usefulness and Td towards novelty,
although there is no formal indication for that on the Model. We are intu-
itively led to this suggestion, and empirical results (in chapter 6) will indeed
confirm this reasoning.
In the imagined scenario given in section 2, we arrived at a point where,
when the system did not have the sufficient knowledge or the search space
was too big to reach a goal in a reasonable amount of time, it would enter a
divergence mode. However, it is clear that the complexity of this mode could
(and possibly would) be higher than the convergent one. In other words, in the
case of our model, if there was a mapping between each pair of the n concepts
from KB, we would have n × (n − 1) bisociations, if the transfer operation,
ω , only produced one new concept for each pair. Since we are assuming a
large and varied knowledge base, this value would be extremely big, even
with an optimistic perspective. Indeed, if we also consider the possibility of
changing T and G , then we soon reach the conclusion that it is unrealistic to
search but only a little portion of this (new) search space. Thus, the purpose
of presenting and discussing this model is not to implement it entirely, but
to state our position on what modelling concept invention with bisociation
is about and to lay the foundations for practical implementations such as the
one we will describe in the next chapter. A question thus arises: Why is this
concept invention?
As we can observe, even if an outcome given by this model is logically de-
ducible from KB, its generation is not based on soundness32 , in the same way
that someone solves a problem without having followed a conscious sequence
of steps. This does not mean that this is the way unconsciousness works, or
even how humans invent concepts, rather it is a model for how it can be com-
putationally simulated. Thus, we call it a model of concept invention because
it produces new (and potentially useful) concepts from an unsound process,
which agrees with the definition we gave in section 2.2.
Again, we would like to raise the question about search. After all, isn’t
this just search in a (complex) space? What more do we have to offer than
any other AI model? The answer is simply yes, it is search. And this is an
AI model which, as many other AI models, aims to simulate a specific kind
of human behavior that has been rarely approached before. In this case, cre-
ativity, more specifically concept invention. Would this mean that creativity
is part of (or is a kind of) intelligence? The answer we give is that they are
100 A Model of Concept Invention

definitely related and, in order to invent a concept, rationality (an indisputable

component of intelligence) is necessary.
Apart from raising philosophical questions, what else is this model for? In
other words, what could be its applications and what is the degree of imple-
mentability of its components? Such a model could be applied in situations
where the generation of new concepts is important, such as in design, archi-
tecture or games, to name a few. Ideally, it could be applied as a meta-level
reasoning engine to help with situations where a lower-level system, ded-
icated to a specific domain, could not find a solution, as suggested in the
imagined scenario described earlier. To some extent any of its modules can
be implemented, however the capability of meta-level reasoning is, perhaps,
the hardest to construct, since it demands self organization and assessment,
two capacities that machines can hardly achieve. For this reason, meta-level
reasoning has not been implemented in our system, Divago, which will be
described next. In this system, we will provide some suggestions for how
other aspects of this model can be implemented, namely cross-space map-
ping, bisociation, the knowledge base, the reasoning engine, the evaluation
and the elaboration.
Chapter 5
Divago

We will now present our system. It is called Divago, after the Portuguese
expression “(Eu) divago”, which means “I wander”. In the previous chapter,
we explained our Model of Concept Invention, which comprehends the main
theoretical substance of this book. In this chapter, we seek to provide a practi-
cal instantiation of its modules. The construction of Divago demanded many
compromises between the overall goal of instantiating the model of concept
invention and the specificities that appeared during the development and re-
flection upon each of the modules. For this reason, there are some points of
conflict between them, namely in the implementation of the search strategies
and in the choice of constraints. Where the reading becomes harder to follow
(due to formalization or algorithms), we will try to synthesize the message in
a manner as fluent as possible.

1. Overview of the architecture

In figure 21, we show the architecture of Divago. Before entering into details,
we prefer to give a superficial overview of how it works, with attention to the
role that each module takes and to the data flow (represented by arrows in the
diagram).
The Knowledge Base contains a set of concepts, each one defined ac-
cording to several different kinds of representations (concept maps, rules,
frames, integrity constraints, instances). The concept maps, rules, frames and
integrity constraints follow the Micro-theory view, while the instances agree
with the Exemplar view.
The first step for the invention of a new concept is the choice of the input
knowledge, in this case a pair of concepts. Since, in Divago, we are focusing
on the mechanisms of divergence and bisociation, we provide no specific al-
gorithm for this selection. This choice is either given by a user or randomly
generated. After being given a pair of concepts, the Mapper builds a struc-
tural alignment between (the definitions of) them. It then passes the resulting
mapping to the Blender, which then produces a set of projections that implic-
itly define the set of all possible blends. This will be the search space for the
102 Divago

Multi-domain Knowledge Base

Constraints
Factory
Blender
Mapper

GA

Convergent Strategy

Elaboration Goal

Figure 21. The architecture of Divago

reasoning mechanism, the Factory.

The Factory is based on a parallel search engine, a genetic algorithm (GA),
which searches for the blend that best complies with the evaluation given by
the Constraints module. Prior to sending each blend to this module, the Fac-
tory sends it to the Elaboration module, where it is subject to the application
of domain or context-dependent knowledge. The GA thus interacts both with
the Constraints and Elaboration modules during search.
The evaluation of a blend given by the Constraints module is based on
an implementation of the Optimality Principles (in section 3.2). Apart from
the blend itself, our implementation of these principles also takes into ac-
count knowledge that comes from the Knowledge Base (namely integrity
constraints and frames), as well as the accomplishment of a goal that comes in
the form of a query. Any of these issues will be described shortly. The Elab-
oration module essentially applies internal-logic elaboration and rule-based
elaboration. The rules involved are also part of the Knowledge Base.
After reaching a satisfactory solution or a specified number of iterations,
the Factory stops the GA and returns its best solution. In some cases, this
result is also the input of an Interpretation module, which produces an in-
terpretation of the new concept. In collaboration with other researchers, we
developed Interpretation modules that generate 2D images (in the house-boat
experiment), textual descriptions (horse-bird experiments) and 3D images
(creatures experiment).
 Knowledge Base 103

Both the Mapper and the Elaboration modules are optional, for different
reasons. The mappings provided by the Mapper are essentially based on the
Analogy and Metaphor works presented in section 4. However, in some situa-
tions, these mappings are very restrictive. Thus, without having implemented
alternative procedures, we allow an externally defined mapping (which, in
some experiments, is user-defined). The Elaboration can also be bypassed for
experimentation reasons. When analyzing results, the elaboration can hide
the real results, i.e. it can fix problems by itself that we may need to watch in
order to assess the functioning of the system.
In comparison with the model presented in the previous chapter, a dif-
ference immediately arises that the mechanisms of divergent and convergent
search (Td and Tc , respectively) are not separated in Divago. On the contrary,
they work intertwined: the method for divergence (the GA) uses the method
of convergence (which applies the Elaboration) once for every blend found.
Another difference is that Divago is not processing its own internal speci-
fications. In other words, we leave meta-level reasoning, which might support
transformational creativity, for future developments. As discussed before, this
is an extremely complex task per se.
We will now describe in greater detail each of the six modules: the Knowl-
edge Base, the Mapper, the Blender, the Factory, the Constraints and the Elab-
oration.

2. Knowledge Base

All representation in the Knowledge Base follows a symbolic approach (as

opposed to sub-symbolic ones, such as neural networks). Nevertheless, we
see no reason to doubt that the same mechanisms could also be applicable
with other representation paradigms. There are many different kinds of struc-
tures in the Knowledge Base of Divago, namely the concept maps, the rules,
the frames, the integrity constraints and the instances. The syntax used in
the Knowledge Base (and in the whole system) is the same as in the Prolog
language.
We call concept maps to the semantic networks that are used in Divago to
describe a concept or a domain. A Concept Map is a graph in which nodes
represent concepts and arcs represent relations. A concept is thus defined in
association to other concepts, which will therefore also intervene within the
concept’s definition. For this reason, some confusion may arise so we ought
104 Divago

to clarify now the notions of domain, concept and element33 . Whenever we

have a concept map in which there is no central concept, rather many different
concepts participate with the same degree of importance, we call it a domain.
Examples of domains could be “biology”, “computers”, “music”. In any of
these, we focus on many concepts rather than only one (e.g. in “music”, we
have “harmony”, “melody”, “rhythm”, etc.). When the concept map focusses
on a single concept x, we say it is the concept map of the concept x. For
example, the concept map of the concept “horse” will have associations to
“nose”, “legs”, “mane”, etc. in order to define what a horse is. Of course, each
of these are also concepts, but as they get farther away from the main concept
(the one about which the concept map is built - in the example, “horse”), they
get less specified (e.g. the concept “human”, from the concept map of “horse”
has no associations to “intelligence”, “face”, “society”, etc. and therefore it
is only superficially identified, possibly with relations such as “owner-of” or
“rider”). In order to avoid confusion, we adopt the convention that each of
these nodes of a concept map will be named element34 , instead of concept.
Thus, a Concept will be made up of the concept maps, rules, frames, etc. as
proposed in the previous chapters, and following essentially the micro-theory
view.
The difference between a domain and a concept is subjective and depends
on the level of granularity. Every domain is by itself a concept and every con-
cept can be seen also as a domain (even if it is a micro-domain). Throughout
this book, unless explicitly stated to the contrary, we assume that a concept
map is defining a single concept, rather than a domain. To sum up, Divago
follows the micro-theory view of concepts (presented in section 1), in which
a concept is defined by facts and rules. We will see that Divago also allows
the use of instances, which agrees with the exemplar view.
The choice of symbols for elements and relations in our concept maps
is arbitrary, yet, mainly after the horse-bird experiment, we followed two
normalization principles. The first one is that relations must belong to (or
descend from) the Generalized Upper Model hierarchy (GUM) (Bateman,
Magnini and Fabris 1995), a general task and domain independent linguis-
tically motivated ontology that intends to significantly simplify the interface
between domain-specific knowledge and general linguistic resources. GUM
occupies a level of abstraction midway between surface linguistic realiza-
tions and conceptual or contextual representations. Being split into two hi-
erarchies, one containing all the concepts and the other all the roles, GUM
gives us a large set of primitive relations to standardize our choices in the
 Knowledge Base 105

concept map. It is important to notice that, in our maps, the members of

the concept hierarchy of GUM (e.g. “color”, “ability”, etc.) are used as
relations (e.g. “color(mane, black)”, “ability(horse, run)”). To know more
about GUM, we redirect the reader to one of the publicly available sources
(http://www.purl.org/net/gum2).
The second principle that we follow in the construction of the concept
maps is that elements in our Knowledge Base may only be represented as
nouns, adjectives, preferably in the singular form, or numerals (in the partic-
ular case of numbers) in English language. As we said, these are only nor-
malization principles for the construction of the concept maps, so, in theory,
the model itself doesn’t take into account the lexical categories of the words
used, following only the principle that “the same word corresponds to the
same element”.
Given the importance that the concept maps have for Divago, we devel-
oped another system, Clouds (which was the subject of an MSc. thesis), with
the goal of helping a user build her own concept maps and avoiding bias as
much as possible. The system led the user towards different areas of the con-
cept map and, as it gets expanded, she can no longer keep track of the whole,
leaving to Clouds the task of leading the construction via questioning the user.
Clouds was used to generate the maps for some of the experiments presented
here. It is far from the theme of this book, so we direct the interested reader
to (Pereira, Oliveira and Cardoso 2000; Pereira and Cardoso 2000; Pereira
2000)).
Let us now define formally a concept map. We will use the same defini-
tions and symbols given in the previous chapter.
Let A E ⊆ A be a set of symbols, which we will call the elements and
let A R ⊆ A be another set of symbols, which we will call the relations. A
Concept Map CM is a set of binary predicates with the form:

X (Y, Z), X ∈ A R,Y, Z ∈ A E

We also define the exhaustive closure CM + as the concept map with all
elements A E and relations A R between them, i.e.

CM + = {X (Y, Z) : X ∈ A R,Y, Z ∈ A E }

Therefore, CM ⊆ CM + .
In tables 6 and 7, we show examples of concept maps for “horse” and
“bird”(made with Clouds). These maps are necessarily arbitrary in the sense
106 Divago

that each person would draw her own maps, a result of the different concep-
tualization and individual points of view one can take. Some relations such as
“pw”(part-whole), “member of” (category inclusion relation), “isa”, “sound”
are either shorter words for the same relations of GUM or extensions made
to this hierarchy.

Table 6. The concept map of horse

isa(horse,equinae) pw(leg, horse) pw(mouth,nose)
isa(equinae,mammal) quantity(hoof, 4) sound(horse, neigh)
existence(horse, farm) pw(hoof, leg) purpose(mouth, eat)
existence(horse, wilderness) pw(nose, horse) purpose(ear, hear)
purpose(horse, traction) eat(horse, grass) color(mane, dark)
motion process(horse,walk) ability(horse, run) size(mane, long)
carrier(horse, human) pw(tail, horse) material(mane, hair)
purpose(leg, stand) quantity(leg, 4) pw(ear, nose)
member of(horse, ruminant) quantity(eye, 2) pw(eye, nose)
purpose(horse, cargo) quantity(ear, 2) ride(human, horse)
purpose(horse, food) purpose(eye, see) pw(mane, horse)
isa(farm, human setting)

Table 7. The concept map of bird

isa(bird, aves) existence(bird, house) isa(aves,oviparous)
lay(oviparous, egg) existence(bird,wilderness) purpose(bird, pet)
purpose(bird, food) purpose(eye, see) pw(beak, bird)
pw(lung, bird) motion process(bird, fly) pw(beak, bird)
pw(wing, bird) quantity(eye, 2) quantity(wing, 2)
isa(owl, bird) isa(paradise bird, bird) quantity(claw, 2)
ability(bird, fly) purpose(lung, breathe) pw(leg, bird)
pw(feathers, bird) smaller than(bird, human) purpose(wing, fly)
purpose(beak, eat) purpose(claw, catch) sound(bird, chirp)
isa(parrot, bird) ability(parrot, speak) pw(straw, nest)
pw(eye, bird) purpose(beak, chirp) purpose(leg, stand)
pw(claw, leg) role playing(bird, freedom) quantity(leg, 2)
isa(nest, container) isa(house, human setting)

We often represent concept maps graphically, in which the relations are

arcs and elements are nodes. Figure 22 shows an excerpt of the concept map
 Knowledge Base 107

Figure 22. Concept map of horse (an excerpt)

for “horse”.
The concept map corresponds to the factual part of the micro-theory of
the concept. The inferential part comprises rules that explain the inherent
causality, frames that have the role of providing a language for abstract or
composite concepts and integrity constraints, particular rules that serve to
assess the consistency of the concept.
Rules have the form:

A0 ∧ A1 ∧ ... ∧ Ai ←− B0 ∧ B1 ∧ ... ∧ B j , Ai ∈ K

with K = CM + ∪ {not R : R ∈ CM + }

This allows for the use of negation as well as a conjunctive set of atoms (Ai ) in
the head, although with specific constraints: Ai ∈ K and B j can be any atom or
logical expression (e.g. a comparison). For example, a rule for inferring that
something is at the gaseous state (and not at the solid or liquid state) could
108 Divago

be:
state(X , gaseous) ∧ not state(X , solid)∧ ←− ebulition lvl(X , N)∧
not state(X , liquid) temperature(X , T )∧
T >N

The syntax followed by Divago for rules is defined by a predicate rule/6:

rule(Domain, Name, PosConds, NegConds, AddList, DelList).

where Domain corresponds to the domain or concept with which the rule is
related, PosConds and NegConds correspond to the (positive and negative,
resp.) sets of conditions of the rule and AddList and DelList correspond to the
(positive and negative, resp.) sets of conclusions of the rule.
The frames have the role of describing abstract concepts, situations or
idiosyncracies. A frame consists of a set of conditions that the concept map
must satisfy. When a concept c satisfies all the conditions of a frame f , we
say that c integrates f . Frames are formally very similar to rules but they are
applied differently in the process:

f rame(Name) : A0 ∧ A1 ∧ ... ∧ Ai ←− B0 ∧ B1 ∧ ... ∧ B j

where Name is an identifier of the frame. A frame should be a meta-level con-

cept that is tightly integrated according to a situation, structure, cause-effect
or any other relation that ties a set of elements onto one, abstract or broad,
concept. For example, the frame of “transport means” (below) corresponds
to a set of elements and relations that, when connected together, represent
something that has a container and a subpart (e.g. an engine) that serves for
locomotion.
f rame(transport means(X )) :
carrier(X , people) ←− have(X , container) ∧ have(X ,Y )∧
purpose(Y, locomotion) ∧ drive( , X )

When a concept map integrates the “transport means” frame, then we can
either say that it is itself a “transport means” or one of its constituents is a
“transport means”. For example, the concept map of school bus would inte-
grate this frame, while the concept map of classroom wouldn’t.
The syntax for representing a frame is the following:

f rame(Domain, Name, PosConds, NegConds, AddList, DelList).

 Knowledge Base 109

An extremely important aspect to remember about frames is that they allow

the inclusion of Prolog scripts inside any of the sets PosConds, NegConds,
AddList and DelList. This offers great power to frames since these scripts
will be run whenever the frame is inspected and integrated. In other words,
in order to check whether a concept map integrates a frame, it will execute
the scripts included in the sets PosConds and NegConds and, during the elab-
oration phase, the frames that are integrated will have their sets AddList and
DelList also executed. We will clarify each of these mechanisms during this
chapter. The frames can thus become externally defined scripts or programs,
executed whenever their conditions apply 35 . We propose to consider several
types of frames according to their degree of abstraction and functional as-
pects. A very specific frame comprehends a well defined set of relations and
elements, such as in “transport means”. A highly abstract frame is one that
considers many different types of elements and relations. For example, when
a concept map satisfies a “noun-noun combination” frame, it means that it
consists of the concept that results from a combination of two nouns - each
one an independent concept in itself (e.g. “pet fish”, “gun wound”).
In terms of their function, frames can be classified as organizing, pattern
identifying or transforming. An organizing frame is a frame that determines
the general structure of a concept map. For example, in the concept map for
school bus, “transport means” could be an organizing frame. A frame is pat-
tern identifying when it matches a pattern within a larger concept map. For
example, in a school trip concept map (which could be a large concept map
with details about many concepts such as bus, teacher, driver, study mate-
rial, theme, route, etc.), one could find a “transport means” frame (focussing
on the school bus part), this time becoming pattern identifying. Normally,
organizing frames are a lot more abstract than the pattern identifying ones,
although sometimes the same frame can take both functions. Transforming
frames may only make sense within a bisociation context (and thus will get
clearer as we progress through this book). A transforming frame identifies a
transformation that occurs during the blending of two input concepts. For ex-
ample, if the new concept map integrates a “new ability” frame, it means that
there were new “ability” elements and relations transferred from one of the
input concepts to the context of the other (e.g. in Pegasus, wings are trans-
ferred from bird to horse, giving it the ability to fly). The specification for
110 Divago

“new ability” could be something like:

f rame(new ability(d1)) :
new ability(X , A) ←− ability(X , A)∧
not rel(d1, ability(X , A))∧
purpose(P, A) ∧ pw(P, X )∧
pro jection(blend, d1, X , X )∧
pro jection(blend, d2, A, A)
Reading this informally, it says that if some element X has, in the concept
map of the blend (that is to say, the new concept map), the ability A, which
was not present in X ’s input space, d1, then we are in face of a “new ability”
given to X . It also says that this “new ability” should have a minimal justifi-
cation, i.e. there must be a subpart P of X whose purpose is to provide ability
A (e.g. if something flies, it should have wings). Furthermore, we can also
require that X and A be projected from different inputs (d1 and d2, resp.) to
the blend.
Thus, a frame will serve, in the process of concept invention, as a tool
for pattern identification, for providing directives for the construction of con-
cepts, and for elaboration. As we will see, frames are essential to control the
system. In Appendix C, we provide a thorough description of these knowl-
edge structures, their specific keywords and some examples.
Although rules and frames are formally very similar, we should now clear
out their distinction and underlying rationales. Rules take an important role
in the definition of a micro-theory of a concept, however in the blend con-
struction, they end up having a passive part (they can be applied to the input
domains before the process, or to the blend, after it is generated). Frames,
on the other hand, can take an active part on the blend construction: they
can be specified as goals by a user and they help structuring the blend, often
becoming the scaffolding around which new blends are constructed. In fact,
as the experiments will show, they are fundamental to help Divago achive
meaningful results.
As the reader may have noticed, our notion of frame departed from mod-
elling Fauconnier and Turner’s “tightly integrated structures” to operational
scripts which actually govern transformations during blend construction. In
cognitive linguistics terms, we believe this partially falls into the image
schema realm (although the definition itself of image schemas in CL has not
reached a consensus, e.g. (Johnson 1987; Neisser 1976; Thomas 1999)) in
the sense that our frames can be patterns that recurrently provide structured
 Knowledge Base 111

understanding of various concepts. Taking an example from Lakoff (Lakoff

1987) of a part-whole schema (in Prolog form):

f rame(generic, pwSchema, [ f indall(X , pw(X ,Y ), L), op(exists(L))],

[ ], [ ], [ ]).

This frame covers every part-whole relation in a concept map (the larger it
becomes, the more “part-whole” based is the concept - the more the schema
becomes meaningful for the concept). It can be said that these frames are too
strict for the “metaphorical” potential of image schemas - the part-wholeness
of the schema should not have to be processed literally (in the same way that
an individual is part of the society, as opposed to a wheel being literally part
of an engine), but it is also true that the set of concepts it can match depends
on the abstractness of the frame. For example, the relation “pw” above could
be replaced by a more general one (e.g. “belonging”) or it could be defined
via an algorithm.
The same reasoning could be taken for any other image schema (once
we find the relationships and their arguments), but it becomes difficult to ar-
gue that frames are image schemata, because we didn’t explore the relation-
ships any further, namely regarding the cognitive basis, the role of perception,
learning or any other aspect regarding the mentality of image schemata. What
we are trying to say is that, with our frames, some of the properties of the im-
age schemata can be simulated, namely being a generic structure, applicable
to different concepts, thus attributing to these concepts a particular associa-
tion (to the image schema - e.g. “part-wholeness”), and triggering associated
inferences. Frames could be seen as a strictly symbolic - computational - ver-
sion of image schemata. Again, we redirect the interested reader to Appendix
C, which can clarify further the scope of these structures.
The integrity constraints serve to specify logical impossibilities. Each
integrity constraint consists of a set of propositions that should not be
simultaneously true. Let Ai be any atom or logical expression, an integrity
constraint ic is defined as:

f alse ← A1 ∧ A2 ∧ ...not An−1 ∧ not An ∧ ...

Two examples of integrity constraints could be for specifying that something

cannot be dead and alive at the same time and for avoiding part-whole recur-
112 Divago

sion, i.e. something cannot have a part-whole relation (pw) with itself:

f alse ← state(X , dead) ∧ state(X , alive)

f alse ← pw(X , X )
In the Prolog syntax that we use, an integrity constraint is represented by the
predicate integrity/3:

integrity(Domain, Pos, Neg)

where Domain is the concept or Domain to which the integrity constraint

belongs and Pos and Negs are the positive and negated conditions of the con-
straint.
Finally, we can define a Concept Micro-Theory (or a Domain theory), CT ,
as being a tuple (CM, R, F, IC), where CM is a Concept Map, R is a set of
rules, F a set of frames and IC a set of integrity constraints.
The micro-theories may be compared to Joseph Goguen’s sign systems
(Goguen 1999). In a sign system, we have sign and data sorts, partial order-
ings on each of these, relations and functions, constructors to build upper
level signs, priorities on these, and axioms. In our micro-theories, we have no
formal distinction between sign and data sort. In principle, every element is
equal, thus its classification and partial ordering can only make sense in a con-
cept map (e.g. an isa ontology with animal classification would correspond
to a sign sort ordering, while another one with colors or numbers would be
a data sort ordering), which also contains the relations and functions. Frames
are our constructors, but there is no ordering or priority over them. Only goal
frames, used in the Constraints module, have priority over the other frames.
Finally, rules and integrity constraints can be seen as Goguen’s axioms.
Another level of representation allowed in Divago is that of the instances.
Along with the micro-theory, one can also add instances to the concept def-
inition (this corresponds to the Exemplar view, as in section 1). Instances
are represented as structures of knowledge that apply (some of) the elements
present in the micro-theory.
Let A A ⊆ A be a set of symbols (the arguments) and let A F ⊆ A be
a set of relations (the functors). Let L C be a set of compositions such that:

x∈AA
c∈LC : c=
I(y1 , y2 , ..., yn ), yi ∈ L C , I ∈ A F
 Knowledge Base 113

Figure 23. An instance of a house

Thus an Instance can be represented in the form I(x1 , x2 , ..., xn ), where I ∈

L F , and xi ∈ L C . xi is compositional, i.e. it can be used as an argument to
another x j . An example of an instance of the concept “house”, as used in the
“house-boat” experiment, could be:
case(1,0, house, [sons=2, size=small, type=simple, son name=roof,
son name=body]).
case(1,0:0, roof, [shape=triangle(30)]).
case(1,0:1, body, [sons=3, in=[left/90,off/25, right/90],son name=structure,
son name=window, son name=door]).
case(1,0:1:0, structure, [shape=square]).
case(1,0:1:1, window, [shape=square(5), in=[off/20, right/90, off/15, left/90]]).
case(1,0:1:2, door, [shape=rectangle(4, 10), in=[off/3]]).

Using the functors ’case’, ’:’, ’=’ and ’/’, this instance describes the several
parts of a house, starting from its top-level element (“house”) to the smallest
constituents (“door”). It associates each of the elements (that are also part of
the theory) to a 2D drawing language (Logo). Its interpretation generates the
image in figure 23.
Since the details of these instances and their syntax are not central for this
book, we redirect the description of the language used to the Appendix D.
Instances are useful for interpreting a new concept in the sense that they
can attach a semantics to a concept and its constituents. For example, with a
visual instance of a house, one sees an example of what a door can look like.
This will be observed in the house-boat and creatures experiments.
Finally, in Divago, a concept is defined by the pair (CT , I), where CT is
the theory and I is a set of instances. The Knowledge Base can simultaneously
have many different concepts, from different domains. However, during con-
114 Divago

cept invention, Divago only considers a pair of concepts (or domains) and a
special domain, the generic domain.
The generic domain contains all knowledge that is applicable to all con-
cepts and to the process of concept invention. It has encoded the hierarchy
of GUM in a predicate arc/5, which can be used to generalize/specialize the
relations found in a concept map. This facility is used by some frames. The
generic domain also contains an isa ontology, which is used mostly by the
Mapper (shown in next section) to build correspondences between elements
of different concepts. The majority of the frames used by Divago also be-
long to the generic domain. In table 8, we show some of the frames from the
generic domain that were used in the experiments.

Table 8. Some frames of the generic space

Frame name Conditions
aframe The blend contains identical structure from input 1
aprojection The blend contains the same elements of input 1
bframe The blend contains identical structure from input 2
bprojection The blend contains the same elements of input 2
new ability An element has an ability relation not existent in its
original input
function transfer An element in the blend has a function that was not
present in its original input.
analogy transfer Transfer all neighbor elements and relations of an
element of one input to the projection of its mapping
counterpart from the other input.

In the generic domain, we find also integrity constraints and rules. An

entire copy of the generic domain is given in Appendix E. In fact, the generic
domain holds most of the knowledge for Divago, while each specific concept
has only encoded the essential, usually consisting solely of the concept map.
To summarize, Divago employs a large variety of knowledge structures,
each one with its own role: concept maps for structural relationships within
a concept (or domain); rules for defining procedures specific to a concept;
frames as abstract concepts that allow the system to identify patterns in con-
cepts and to infer further knowledge (running the blend); integrity constraints
to establish limits for reasoning; and instances, as examples of the concept
(or domain). As we have said, these do not all have to cohabit simultaneously
in the system. Indeed, with concept maps, frames and integrity constraints,
 Mapper 115

the system is able to give results as shown in chapter 6. The need of each
representation type will depend on the problem at hand: if working with con-
cepts that need reasoning that depends on “hidden” inferences (e.g. temporal
reasoning - story plot blending; spatial reasoning - scenario blending), we
will need rules; if we intend to design specific objects (e.g. 3D objects), then
instances may be necessary to provide a real output; if making experiments
at the conceptual level (e.g. testing examples of Conceptual Blending), then
maybe frames and concept maps will be enough. The integrity constraints
are present in almost every application, as they become essential to advise
Divago not to follow wrong paths36 .

3. Mapper

The Mapper defines the mapping function φ of the model presented in this
book. In the definition we gave, this function is oversimplified, since it only
provides a binary association between pairs of concepts. In Divago, a con-
cept is itself a structure with many different sub-structures, thus the mapping
becomes somewhat more complex. This justifies a revised version of the map-
ping operation, φ :
φ : U × U −→ M
where M is the powerset (set of all sets) of all possible pairs of sub-structures
from concepts of U . In this module, we propose to use exclusively the con-
cept maps. To state this more clearly, for any pair, CM1 and CM2 , of concept
maps, the Mapper will find a set of mappings between their elements, each
pair having one element from CM1 and one from CM2 .
In his formalization of Conceptual Blending, Goguen introduces the semi-
otic morphisms. A semiotic morphism is a structure preserving mapping, as
it should map sorts to sorts, subsorts to subsorts, data sorts to data sorts,
constructors to constructors, etc. (Goguen 1999). However it is assumed that
these should only be partial maps. As far as we are aware, Goguen does not
suggest any specific algorithm for building semiotic morphisms. We believe
that an algorithm such as the one implemented for the Mapper could be a
viable solution.
The Mapper uses a spreading activation algorithm to look for the largest
isomorphic pair of subgraphs (contained in the concept maps). In this con-
text, two graphs are considered isomorphic when they have the same rela-
tional (arcs) structure, independently of the elements (nodes). There is poten-
116 Divago

tially more than one structure matching between any pair of concept maps
and this complexity grows worse than exponential with the number of ele-
ments (nodes)37 . However, since it only allows alignment when it finds equal
relations in both graphs, the number of possible solutions can be drastically
reduced, although still demanding that Mapper makes the search in a huge
space. Furthermore, the algorithm starts with a randomly selected pair of el-
ements, so the “perfect choice” (or even the same choice) is not guaranteed
every time we run it.
The Mapper uses an algorithm of structure matching inspired by Tony
Veale’s Sapper framework (Veale 1995). We have already presented Sapper
and therefore the differences will now be further enhanced. While Sapper
needs two cycles to obtain the mapping (one for laying down dormant bridges
with the triangulation rule and one for finding the mapping)38 , our Mapper
uses three cycles: one for laying down dormant bridges (with both triangula-
tion and squaring rules); another one for spreading activation (in our case, a
flood fill algorithm); and a final cycle for finding the mapping. In the first cy-
cle, Mapper builds new dormant bridges whenever two elements from the two
input concept maps share the same relation to the same element (the triangu-
lation rule). Here, the generic domain (and particularly the isa-ontology) is
extremely important because it is a source of shared knowledge. The Mapper
also adds a dormant bridge between every two elements that share the same
relation to two different elements that are connected by a dormant bridge
(the squaring rule). Thus, while Sapper adds dormant bridges as the mapping
is found, the Mapper creates all possible dormant bridges in the first cycle.
The second cycle in Mapper spreads activation throughout the concept maps.
This is different to Sapper where this activation has no prime factorization
or wave. It has only an activation value that decays as it passes by elements.
This activation starts at 100 and is reinforced when passing near a dormant
bridge. Below a threshold (the default value is 20), it stops spreading. After
this second cycle, the network will have a set of sub-graphs with activated
elements, centered in the dormant bridges. The final cycle starts with the ran-
dom (or user-given) choice of one of the dormant bridges, the seed mapping.
This dormant bridge is awakened, and thus becomes the first mapping. Then
it progresses in parallel in both concept maps, so that each new pair of ele-
ments to be mapped (i.e. each dormant bridge visited) is connected by a pair
of equivalent relations to a previously awakened dormant bridge.
As a result of the algorithm, the Mapper returns a set of mappings between
the two concept maps. This module was born out of an idea to implement a
 Blender 117

version of Sapper that would not worry about returning the best (widest?)
mapping or would bias the mapping towards the highly activated nodes (for
instance, in Sapper, the choice “Scalpel: Snub-Fighter” beats out “Scalpel: B-
52” (Veale 1995: chapt. 6) due to higher activation; in Mapper, any could be
selected). The principle was that, if it clearly became less effective (slower,
with smaller mappings, etc.) than Sapper, then we would directly use Veale’s
algorithm. However, we gradually found that the Mapper had limitations that
would not be resolved by changing to Sapper. As we will see in the exper-
iments, restricting mappings to structure alignment narrows the potential of
the system, thus we gave the Mapper a secondary (i.e. optional) role in Di-
vago. On the other hand, the behavior of the module was sufficiently satis-
factory to be retained in some situations. In spite of the complexity involved,
this module is fast in returning a mapping and it achieves the same results as
Sapper in the majority of the time .
As an example, we show in figure 9 the three different mappings produced
for the concept maps of horse and bird (from tables 6 and 7). It is important
to understand that every relation has the same weight in the graph and there
is no domain knowledge or special heuristics considered in the mapping con-
struction. This means that the results may contain non-intuitive associations
(e.g. “4” associated with “2”; “nose” with “bird”).

4. Blender

In Goguen’s algebraic semiotics Blending formalization (Goguen 1999), a

blend is some sign system that results from the semiotic morphisms from the
input and generic spaces (the injections). These morphisms should be mutu-
ally consistent. The “best blend” (to what he calls 3/2 pushout) would thus
result in an ordering of semiotic morphisms by quality, e.g. they should be as
defined as possible, should preserve as many axioms as possible, and should
be as inclusive as possible (i.e. contain the maximum number of mappings
between concepts). Although this author considers this “best blend” as the
best result over a conjunction of compromises between criteria, again it is not
clear what exactly these criteria are, apart from structure mapping. Generi-
cally, we follow some of the same ideas as Goguen, namely the application
of criteria for ordering a set of candidate blends, and having in structure a
fundamental index for quality. The Blender is the first part of this approach
(which will be completed with the Factory, the Constraints and the Elabora-
118 Divago

Table 9. The three mappings

vegetable food ↔ vegetable
food ↔ food
ear ↔ wing horse ↔ bird
nose ↔ bird equinae ↔ aves
eye ↔ lung animal ↔ animal
mouth ↔ feathers human setting ↔ house
2 ↔ 2 wilderness ↔ wilderness
hear ↔ fly ruminant ↔ oviparous
run ↔ fly
1 cargo ↔ pet
neigh ↔ chirp
nose ↔ lung
mane ↔ feathers
tail ↔ beak
mouth ↔ beak leg ↔ eye
nose ↔ bird hoof ↔ wing
eye ↔ lung 4 ↔ 2
ear ↔ feathers eye ↔ leg
eat ↔ eat ear ↔ claw
hear ↔ catch
2 grass ↔ grass

tion module) and focuses on calculating the set of all possible blends.
Assuming a mapping m, generated by the mapping operation φ , as de-
fined by the Mapper or by an external source, we must specify the transfer
operation, ω , which will transfer knowledge from two concepts into one (as
in chapter 4). As with the mapping function, so the transfer operation works
with the concept maps.
First, we have to define what a blending projection is. A blending projec-
tion of an element x from concept map CM is a non-deterministic operation
that maps x to another element (in the blend) which is either x, Ø, x|y or y (y
is the counterpart of x, i.e. (x, y) ∈ m). The symbol x|y is called a compound
and can be read as being both x and y at the same time. In order to consider
this symbol, we must have the alphabet AB , which contains the alphabet A
 Blender 119

plus every combination of pairs x|y that are possible to obtain from symbols
of A . Thus, given A and AB , two concept maps CM1 and CM2 (the two input
concepts), a mapping m (given by φ ), a blending projection γ is the operation
γ : A −→ AB , such that:
⎧
⎪
⎪ x ∨ x|y ∨ y ∨ Ø if x ∈ CM1 , ∃y ∈ CM2 : (x, y) ∈ m
⎨
x ∨ y|x ∨ y ∨ Ø if x ∈ CM2 , ∃y ∈ CM1 : (y, x) ∈ m
γ (x) =
⎪
⎪ x∨Ø if (x ∈ CM1 , ∃y ∈ CM2 : (x, y) ∈ m) or
⎩
(x ∈ CM2 , ∃y ∈ CM1 : (y, x) ∈ m)

Figure 24. The blending projection applied to two small concept maps

Informally, a blending projection determines, for each element of a con-

cept map, what its correspondent will be in the blend. When such an element
(x) has a counterpart in the mapping, then it can be projected as a copy (x), as
120 Divago

a compound with the counterpart (x|y), directly as its counterpart (y) or have
no projection at all. For example, from the third mapping in figure 9, “wing”
could be projected to “wing”, “hoof|wing”, “hoof” or be absent in the blend.
In figure 24, we sketch all possible projections from two little concept maps.
Notice that “human” has no mapping counterpart, therefore it can only map
to its copy or to Ø.
A blend is defined by the blending projections. The transfer operation, ω ,
is defined by an algorithm that composes the blend by transferring knowledge
from the inputs to the blend, according to the projections. It corresponds to the
step of Composition of the Conceptual Blending framework. The algorithm
follows:

Input:
Two input concepts, C1 and C2 , defined by the pairs (CT1 , I1 ) and (CT2 ,
I2 ), respectively, with CT1 =(CM1 , R1 , F1 , IC1 ) and CT2 =(CM2 , R2 , F2 ,
IC2 ).

Algorithm:
Let Blend ←− {}
For i=1,2 do
For each relation r(a, b) in concept map CMi do
Add relation to Blend with the form
r(γ (a), γ (b)), iff γ (a) and γ (b) are not Ø
EndDo
For each rule r from Ri , in the form
r = c1 (x1 , y1 ) ∨ c2 (x2 , y2 ) ∨ . . . ∨ cm (xm , ym ) ∨ {Codec } ←− p1 (z1 ,t1 )
∧p2 (z2 ,t2 ) ∧ . . . ∧ pn (zn ,tn ) ∧ {Code p }, do
Add new rule to Blend such that each cm (xm , ym ) is substituted by
cm (γ (xm ), γ (ym )) (when γ (xm ), γ (ym ) = Ø) and pm (zm ,tm )
is substituted by pm (γ (zm ), γ (tm )) (when γ (xm ), γ (ym ) = Ø)
Copy Code p and Codec (the scripts) directly to the new rule.
EndDo
For each frame f from Fi do
Apply the same process as with rules
EndDo
For each integrity constraint ic from ICi do
Apply the same process as with rules
EndDo
For each instance s from Ii , in the form s = I(x1 , x2 , ..., xn ), do
Add new instance to Blend with the functor I and apply the
same process as with relations (but with arity n and recursively)
EndDo
EndDo
 Blender 121

This algorithm basically creates a blend by applying the projections to

all the constituents of the input concepts. As the projection operation, γ , is
non-deterministic, when this algorithm is applied without selecting specific
projections (i.e. without restricting to only one projection for each element of
the concept maps, as will be done by the Factory module), it does not produce
a single blend, rather it generates what we call a blendoid. The blendoid con-
tains all possible constituents (relations, rules, frames, instances and integrity
constraints) that can be present in any blend of two specific input concepts. In
other words, it implicitly includes all the search space of blends (see example
in figure 25).

M bird
horse
pw pw
hoof
ride M wing

human

Input 1 Input 2
(horse) (bird)

horse|bird
horse

bird
ride
ride pw
ride pw pw
pw

pw hoof|wing pw
human pw
pw pw
hoof
wing
blendoid

Figure 25. A blendoid

Taking a close look over the search space of blends, we notice that, for an
input concept 1 with a concept map with m different elements and  an input

concept 2 with n elements, we may have the maximum of ∑ki=1 [ mi · ni · i!]
different mappings (if we use the isomorphic mappings, as in the Mapper),
122 Divago

with the largest mapping having a size k = min(m, n). To understand this for-
mula, let us consider the extreme case in which all relations (in both concept
maps) are equal. Counting all mapping sizes (from 1 to k), we have, for a
mapping size39 i, all combinations of m elements matching all arrangements
of n elements. In reality, the number of mappings is much lower since there
is a variety of different relations in both inputs. Furthermore, we may also
assume that the Mapper will normally produce only the largest mappings
(smaller mappings are generated only when the Mapper loses activation pre-
maturely when doing the spreading activation process)40 .
Assuming each blending projection is independent, we will have a total
of l = m + n different projections in the definition of every blend. So, in the
“least complexity scenario”, the size of the mapping is 0, meaning that we
only have two choices for each of the l elements (either it gets projected to
the blend or it is not projected), thus we have 2l different blends. If the size of
the mapping is k (the maximum possible), we have four choices for each of 2k
elements (k elements in each of the domains) because each element x mapped
to y can be projected either to x, y, x|y or Ø. Apart from these 2k elements, the
rest (l − 2k) has only two possibilities. This leads us to the conclusion that, for
a mapping of size s (0 ≤ s ≤ k), we have 42s × 2l−2s different possible blends
to choose, which is a very large search space. For example, for m = n = 20
(an “average” sized pair of networks), we have at least 240 (if s = 0) and at
most 420 (if s = 20) different solutions. If we remember that the Optimality
Principles are mutually competing pressures, then we may guess that this is
a very complex search space. Obtaining a good blend is the main motivation
for the Factory module, which will be the subject of the next section.
The Blender module provides two fundamental services to the Factory: it
generates the blending projections (only once and before the Factory starts
searching); it provides the transfer operation, which is used by the Factory
each time it needs to create a blend.

5. Factory

The Factory is the processing core of Divago. It corresponds to the reasoning

mechanism of our model of concept invention and is responsible for applying
the divergent strategy, Td , which is, as we will see, encoded as a genetic
algorithm.
In our context, the output of the Factory, i.e. the invented concept, will
 Factory 123

Figure 26. A possible selective projection from the example of figure 24

correspond to a blend. Since a blend is primarily defined by a string of pro-

jections, then searching for an invented concept becomes searching for the
string of projections that originates the best blend. A string of projections
that comprises one and only one projection for each element from the input
concepts is called a selective projection. Therefore, as discussed in the pre-
vious section, for a pair of input concept maps, and a mapping of size s, we
have a number of 42s × 2l−2s different possible selective projections.
Assuming γ (x) as an algorithmic function that returns each time one and
only one of the possible projections allowed by γ (x), and the pair of concepts
C1 and C2 (containing concept maps CM1 and CM2 , respectively), a selective
projection, λ can be defined as:

λ (C1 ,C2 ) = {γ (x1 ), γ (x2 ), . . . , γ (xm ), γ (y1 ), γ (y2 ), . . . , γ (yn )},

with xm ∈ A E CM1 ∧ yn ∈ A E CM2

with A E CMi corresponding to the set of elements that are present in CMi .
Given the complexity of the search space, and for computational reasons,
we decided to implement a parallel search algorithm, a genetic algorithm
124 Divago

(GA): a framework inspired by evolutionary theories in which we have a se-

quence of populations of individuals, each individual having a fitness value
that represents its survival and reproduction possibilities. Each individual has
a genotype, its birth given genetic code, and a phenotype, the actual interpre-
tation of the genotype (in nature, the animal itself). A genetic algorithm works
as follows:

1. N individual genotypes are randomly created (sometimes the re-

searcher might have an idea as to what is a good genotype and would
direct the creation of the initial population). These individuals are the
initial population.

2. Each individual in the population is evaluated by a fitness function.

This evaluation is based on the phenotype.

3. The best individuals are chosen for reproduction. This choice can be
based on aspects other than fitness value (e.g. biodiversity).

4. The genotypes of the chosen individuals reproduce using the methods

of crossover and mutation and a new population is formed. Other op-
erators are also used. In our case, we also use asexual reproduction
(direct copy of the genotype)

5. Steps 2 - 4 are repeated for a set amount of times or until a halt condi-
tion is met.

This well-known framework has had much success in problems with a

search space with the complexity characteristics that we described. Further
explanation of GA’s is far outside of the scope of this book, so we direct the
interested reader to (GAs:Goldberg 1989).
In our GA, the genotype corresponds to a “selective projection”. The indi-
vidual is thus an ordered sequence of projections (the genes), each one with
one of the allowed values (from the set x, y, x|y and Ø). The phenotype is
constructed with the transfer operation, as given by the Blender module, and
elaborated by the Elaboration module. In figures 27 and 28, we show exam-
ples of the genotype and the phenotype (generated with the transfer opera-
tion).
 Factory 125

Figure 27. The genotype of the individual corresponding to the selective projection
of figure 26

Figure 28. The phenotype - a horse|bird has wings and is ridden by humans

The initial population has 100 individuals selected randomly. The evalua-
tion of an individual is made by the application of the Optimality Principles,
which then participate in a weighted sum, yielding the fitness value. This
work is performed in the Constraints module. The selection of the individ-
uals is of the roulette-wheel type, i.e. the ones with higher fitness have a
greater probability of being chosen. Our algorithm uses 3 reproduction oper-
ations: asexual reproduction (the individual is copied to the next population);
crossover (two individuals exchange part of their list of projections) and mu-
tation (random changes in the projections). It also allows the random genera-
tion of a new individual, which can be useful when the current population has
low biodiversity. The system stops when a predefined number of iterations of
this process has been done, when it stabilized around an acceptable maxi-
mum for more than a predefined number of iterations or when an individual
was found that has a satisfactory value.
126 Divago

As we have said before, the choice for a GA is connected to computa-

tional preferences, thus we do not make a claim for any cognitive implica-
tions in this matter. Other choices could be followed (e.g. simulated anneal-
ing), but GAs are highly versatile (with a simple tweak in the process, it can
be changed into a simulated annealing) and offer a bulk of experimental and
theoretical bases from which to apply methodologies or choose parameters.
With this GA, Divago is able to search in a huge space of blends according
to the preferences of the user. The best solution is not guaranteed, but it is
reasonable to expect that the higher the number of iterations, the more likely
it is to find a good blend, if one exists in the search space.

6. Constraints

The Constraints module implements the Optimality Principles. It makes a

preprocessing of each blend (checking frame satisfaction and completion, in-
tegrity constraint violation, vital relation projection, etc.) and then obtains
a value for each of the eight measures. These values then participate in a
weighted sum, which yields the value of the blend (normalized to fall into the
[0,1] interval) which is returned to the Factory. The weight attributed to each
optimality pressure is defined by the user. Our proposal for a computational
realization of the eight Optimality Principles concerns solely the representa-
tion and scope of this model41 . This doesn’t mean that this proposal should
not be verified or tested with regard to cognition and the blending phenom-
ena in general, rather we state that we didn’t base our measures on cognitive
experiments, but mainly tried to follow the philosophy behind the description
given by Fauconnier and Turner in the context of our formal model.
While these constraints consider usefulness, as well as many other aspects,
it is clear that they lack any explicit concern to novelty. Instead of adding an
extra constraint for novelty, we decided to face it as an effect rather than as a
construction principle. In other words, we intend to verify if, with the present
architecture and constraints, Divago can produce novelty. In the experiments,
we will be able to see that it is capable of some degree of novelty.
We will now present our computational version of the eight optimality
principles.
 Constraints 127

Figure 29. The blend satisfies (or accomplishes) the frames “Equinae” and “Flying
thing”

6.1. Integration

Frames gather knowledge around abstractions, tightening the links between

elements. They organize a concept into a more understandable whole. For
example, in figure 29, two specific frames integrate the blend into a more
broad concept of “flying equinae”.
Assuming the set F of frames that are satisfied in a blend, we define the
frame coverage of a blend to be the set of relations from its concept map
that belong to the set of conditions of one or more frames in F. The larger
the frame coverage of the blend, the higher the integration value should be.
Yet, a blend that is covered by many frames should be less integrated than a
blend with the same coverage, but with less frames. In other words, if a single
frame covers all the relations of a blend, it should be valued with the maxi-
mal integration, whereas if it has different frames being satisfied and covering
different sets of relations, it should be considered less integrated. The intu-
ition behind this is that the unity around an integrating concept (the frame)
reflects the unity of the blend. The Integration measure that we propose varies
according to this idea. It also takes integrity constraints into account so that,
when a frame violates such a constraint, it is subject to penalty.
Definition 6.1 (SingleFrameIntegration). For a frame f with a set C of con-
ditions, a blend b, with a concept map CMb , its blendoid with a concept map,
128 Divago

CMB+ , and V I, the set of integrity constraints that are violated in the frame,
the integration value, I f is defined by:
#C #CMb
If = × (1 − ι )#VI × (1 + )/2
#CMb #CMB+
being ι a penalty factor between 0 and 1, a value that penalizes a frame for
each violation of integrity constraints. An integrity constraint is violated if its
premises are true. In the context of the integration measure of frame f above,
f violates integrity ic if the conditions Cic of ic are verified and Cic C = Ø.
In other words, f needs to violate ic in order to be integrated.
We would like to further clarify the above formula: the first factor repre-
sents the ratio of coverage of b w.r.t. f ; the second factor means that each
integrity constraint violation implies an exponential discount; the third factor
serves the purpose of maximizing the size of the blend (if two frames have
the same ratio of coverage, the one that contains more relations should have
higher integration); the division by 2 aims to normalize the result between 0
and 1.
While the value for a single frame integration is described above, the in-
tegration measure of a blend w.r.t. a set of frames is not necessarily straight-
forward. At first sight, it is appealing to just add the values of integration of
all frames, or of the union of their condition sets. Or even their intersection.
But this would lead to the wrong results, because a set of frames can not be
reduced to a single frame from the point of view of integration. In this mea-
sure, we want to stimulate unity, coverage and take into account the strength
of each frame individually. In terms of unity, we argue that the set of relations
that make the “core” of all the frames that are satisfied (i.e. the intersection of
the sets C of conditions of all frames) should be highly valued. On the other
side, the coverage of this “core” will be smaller than the overall coverage
(or equal, if the frames have equivalent C sets), which leads us to take into
account the disjoint sets of relations of the frames. Finally, the integration of
each individual frame (as defined above) should also be present in the overall
measure. These last two issues (the overall coverage and the integration of in-
dividual frames) are subject to a disintegration factor because they reflect the
existence of different, not totally intersected, frames. We propose this factor,
α , to be a configurable value from the interval [0, 1]. It is now time to present
our proposal for the Integration measure of a blend:
Definition 6.2 (Integration). Let Fb = { f1 , f2 , ..., fi } be the set of the frames
that have their conditions (Ci ) satisfied in the blend b, α , the disintegration
 Constraints 129

factor (with 0 < α < 1), and I fi , the single frame integration value, as in 6.1.

i
Integration = I i + α ×U ncoverage × ∑ I fi
0 Ci
0

The U ncoverage value consists of the ratio of relations that do not belong
to the intersection of all frames w.r.t. the total number of relations considered
in the frames:
i i
# 0 Ci − # 0 Ci
U ncoverage =
# i0 Ci

The integration measure is fundamental to the blending process. It leads

the choice of the blend to something recognizable as a whole, fitting patterns
that help to determine and understand what a new concept is.
In order to illustrate this reasoning, in figure 30, we show 4 blends and
the respective frame coverage. Blend A clearly gets the highest Integration
value (all the relations are covered by a single frame); B is also totally cov-
ered, but by two different frames; Blend C should get lower Integration value
than B because it does not cover every relation (Uncoverage is higher than
0); finally, blend D would possibly get the lowest value (depending on the
value of α ) because, although covering every relation, there is a high disper-
sion of frames. To help make the calculations clearer, let us pick specifically
the blends A and D and determine their Integration values. To simplify, let us
assume that CMB+ = CMb (the blendoid size is only necessary for normaliza-
tion purposes, so any size would do for this example) and that there are no
integrity constraint violations. Our blend A will have a single frame integra-
tion, I f , of 1, because it contains all 12 relations of the blend. Since it is the
only one and there are no uncovered relations (U ncoverage is 0), the overall
Integration will also be 1. Now, for the Integration of the blend D, we must
2 2 2
repeat the process. There will be four frames (with I f values of 12 , 12 , 12 and
1
12 ). There is no intersection between all frames, which implies that the first
term of the sum will be 0. Furthermore, no two relations belong to different
frames (U ncoverage is maximum, i.e. 1) and the sum of our four single frame
integrations will be 12 2
+ 122
+ 12
2
+ 12
1
= 127
. This leads us to the second term
of the Integration measure: α × 1 × 12 (remember α is smaller or equal to 1).
7

This shows that the Integration value of blend D will be considerably smaller
than that of the blend A, as intended.
130 Divago

Figure 30. The role of frame coverage in Integration value

6.2. Topology

The Topology optimality pressure brings inertia to the blending process. It

is the constraint that drives against change in the concepts. This happens be-
cause, in order to maintain the same topological configuration as in the inputs,
the blend needs to maintain exactly the same neighborhood relationships be-
tween every element, ending up being a projected copy of the inputs. In prac-
tice, this pressure is normally one that is disrespected without a big loss in
the value of the blend. This is due to the imagination context that normally
involves blends, i.e. novel associations are more tolerable. Of course, this still
depends on the type of blend we are pursuing: if it regards an analogical or
structure alignment construction (e.g. the Buddhist Monk), then Topology is
vital; if it regards a free combination (e.g. a “horse-bird”, an imaginary object
with a goal), then Topology may become secondary.
In our Topology measure, we follow the principle that, if a pair of el-
ements, x and y, is associated in the blend by a relation r, then the same
relation must exist in one of the inputs. If so happens, we say that r(x, y) is
topologically correct. Thus, the value of Topology corresponds to the ratio of
topologically correct relations in the concept map of the blend.

Definition 6.3 (Topology). For a set TC ⊆ CMb of topologically correct

relations, such that

TC = {r(x, y) : r(x, y) ∈ CM1 ∪CM2 )}

where CM1 and CM2 correspond to the concept maps of inputs 1 and 2, re-
spectively, the topology measure is calculated by the ratio:

#TC
Topology =
#CMb
 Constraints 131

Intuitively, this measure represents the amount of relations from the in-
puts that were projected unchanged to the blend. We are aware that this
is stricter than the topology constraint in the blending framework, as it is
based on identity rather than counterparts. As the reader will see, this role of
analysing counterparts as maintaining original structure/neighborhood is left
for the Unpacking constraint. At the moment, the only way to violate topol-
ogy is by having a pair of concepts projected to the same one (e.g. “horse”
and “bird” projected to “horse”), bringing a new relation that was exclusive
to one of the domains (e.g. ability(bird, f ly) projects to ability(horse, f ly);
pw(wing, bird) projects to pw(wing, horse)).

6.3. Pattern Completion

The Pattern Completion pressure brings the influence of patterns either

present in the inputs or coming from the generic space. Sometimes a concept
may seem incomplete but makes sense when “matched against” a pattern.
At present, in the context of this work, a pattern is described by a frame,
i.e. we don’t distinguish between these two notions, and therefore pattern
completion is basically frame completion. The act of completing a frame con-
sists in asserting the truth of the ungrounded premises (which is done in the
Elaboration module), a process that happens only after a sufficient number of
premises is true. We call this the completion threshold, a value that is exter-
nally configured in Divago. To the measure regarding the conditions that are
actually satisfied by a frame f in a blend b, we call the completion evidence
of f , e( fi , b). Therefore, completion can only happen when the completion
evidence is higher than the completion threshold.
Definition 6.4 (Completion Evidence). The completion evidence e of a
frame fi with regard to a blend b is calculated according to the following.
#Sati
e( fi , b) = × (1 − ι )#V I
#Ci
where Sati contains the conditions of each fi that are satisfied in b, ι is the
integrity constraint violation factor and V I the set of violated integrity con-
straints.
As in the Integration constraint, we have the problem of taking into ac-
count multiple frames. This time, given that we are evaluating possible com-
pletion of subsets of relations, instead of sets of relations that are actually
132 Divago

Figure 31. Pattern Completion examples

verified in the blend, it is difficult to find such a linear rationale (e.g. would
two patterns each with an individual completion x value higher than three
each having slightly less than x?). As a result, we propose to find the union
of all the conditions contained within the patterns and then estimate its own
completion evidence:
Definition 6.5 (Pattern Completion). The Pattern Completion measure of a
blend b with regard to a set F with n frames is calculated by
PatternCompletion = e(∪n0 fi , b)
This measure has a very important role in increasing the potential of
the blend, for it brings the “seeds” that may be used in the Elaboration
module. In figure 31, we illustrate Pattern Completion with two examples.
Assuming a frame with three conditions (pw(X ,Y ), purpose(Y, f ly) and
ability(X , f ly)), on the left it has a completion evidence of 66.6% (two re-
lations out of three are already accomplished: pw(Horse Bird, wings) and
purpose(wings, f ly)), whereas on the right the completion evidence is only
33.3% (only pw(Horse Bird, wings) exists). For both, since we consider only
one frame (i.e. one pattern), the value of Pattern Completion is the same as
of the completion evidence.

6.4. Maximization/Intensification of Vital Relations

Fauconnier and Turner propose a set of vital relations that should govern the
blend creation (Fauconnier and Turner 2002). As already stated, our approach
to CB relies essentially on earlier works from the same authors, thus we must
point out that the approach to vital relations discussed here leaves more open
questions that encountered solutions. We may say that we are facing vital re-
lations as merely salient relations, without agonizing with their role in com-
pression - which should be, one might say, their essential role. Compression
 Constraints 133

is the phenomenon of bringing relations between concepts from different in-

puts (the outer-space relations) to the blend (i.e. they become inner-space
relations in the blend)42 . Given the fuzzy definition of Intensification of Vital
Relations and the fact that we are not focussing on compression for this work,
the distinction between Maximization and Intensification becomes yet an-
other step towards subjectivity, which we intend to avoid. In practical terms,
this means we only propose here a measure for Maximization and leave the
discussion of the distinction and specific measure of Intensification for further
research (possibly when a computational view of compression is explored in
depth)43 . By default, we allow the same vital relations44 between and within
two concept maps, some being only rarely used (e.g. change, disanalogy, in-
tentionality). Divago also accepts the definition of other relations as being
vital. For example, in inventing concepts for a game, one may decide that
the vital relations are “strength”, “defense”, “ability” and so on. The effect
of this choice may be that, when giving more emphasis (higher weight in the
fitness function) to Maximization of Vital Relations, the resulting blends will
contain the maximum possible number of these relations.
For implementing this measure, we estimate the impact of the vital rela-
tions to the blend calculated by the ratio of vital relations w.r.t. the whole set
of possible vital relations, contained within the blendoid.

Definition 6.6 (Maximization V R). Let ϒ be a set of vital relations. From the
concept map of the blend b, we may obtain the set of vital relations in b, BV R :

BV R = {r(x, y) : r(x, y) ∈ CMb ∧ r ∈ ϒ}

From the blendoid (the union of all possible blend), B+ , we have BV+R :

BV+R = {r(x, y) : r(x, y) ∈ CMB+ ∧ r ∈ ϒ}

Finally, the Maximization of Vital Relations measure is calculated by the ratio

#BV R
Maximization V R =
#BV+R

6.5. Unpacking

Unpacking is the ability to reconstruct the whole process starting from the
blend. In our view, such achievement underlies the ability to reconstruct the
134 Divago

input spaces. I.e. the reconstruction of the input spaces from the blend de-
mands the assessment of the cross-space mappings, the generic space and
other connections. Thus, what we are proposing is that Unpacking can be re-
duced to the ability to reconstruct the inputs. This is because there is no way
to properly reconstruct the inputs without a reconstruction of the cross-space
mappings, generic space and the connections between spaces.
Unpacking should take the point of view of the “blend reader”, i.e. some-
one or something that is not aware of the process of generation, thus not
having access to the actual projections. Being such, this “reader” will look
for patterns that point to the “original” concepts. Once again we use the idea
of frames, more specifically the defining frame of an element, which com-
prises the immediate surrounding elements and relations. For example, if the
element “wing” was projected onto x in the blend, the defining frame with
regard to the “bird” concept map would consist of purpose(x, fly), condi-
tional(x, fly), quantity(x, 2) and pw(x, bird). The more that these relations are
found in the blend, the more likely it is that the “reader” will find it easy to
understand the relationship between x and “wing”.

Definition 6.7 (De f iningFrame). Given a blend b and an input space d, the
element x (which is the projection of the element xd of input concept map d
to b) has a defining frame fx,d consisting of

fx,d = C0 ,C1 . . .Cn −→ true

where Ci ∈ {r(x, y) : r(xd , y) ∈ CMd }. Assuming that k is the number of con-
ditions (Ci ) of fx,d that are satisfied in the blend, the unpacking value of x
with regard to d (represented as ξ (x, d)) is

k
ξ (x, d) =
n
where n is the number of elements to which x is connected. We calculate
the total estimated unpacking value of x as being the average of the unpacking
values with regard to the input spaces. Thus, having input concept maps 1 and
2, we have
ξ (x, 1) + ξ (x, 2)
ξ (x) =
2
Definition 6.8 (U npacking). Let X be the set of m elements of the blend
b, generated from input concept maps 1 and 2. The Unpacking value of b is
 Constraints 135

Figure 32. Unpacking examples

calculated by
i=0 ξ (xi )
∑m
U npacking = , xi ∈ X
m
In Figure 32, we present the defining frame for “horse”, in the “horse”
concept map. In Blend 1, the element “horse|bird” (the projection of “horse”)
will have the highest Unpacking value (w.r.t. “horse” concept map) because
it fits precisely into its defining frame. In Blend 2, the value is lower because
there are two new relations (with “fly” and “wings”), meaning that it is not
exactly same element. Blend 3 will get the lowest Unpacking value of all
three because it also lacks some relations (e.g. with “run” and “grass”).

6.6. Web

The Web principle concerns being able to “run” the blend without cutting the
connections to the inputs. In our opinion, this is not an independent princi-
ple, being co-related to those of Topology and Unpacking because the former
brings a straightforward way to “maintain the web of appropriate connections
to the input spaces easily and without additional surveillance or computation”
and the latter measures exactly the work needed to reconstruct the inputs from
the blend. This is not to say that Web is the same as Topology or Unpacking.
Rather, on one side, Topology provides a pressure to maintain the most fun-
damental connection to the input: the same structure; on the other side, Un-
packing evaluates the easiness of reestablishing the links to the inputs. These
two values combined in a weighted sum yield, we propose, an estimation of
the strength of the web of connections to the inputs:
Definition 6.9 (Web).

Web = λ × Topology + β ×U npacking

136 Divago

with λ + β = 1.

Since this is not an independent variable, making independent experiments

with the Web measure would not add any valuable conclusion and thus it will
not be covered in this book.

6.7. Relevance

The notion of “relevance” or “good reason” for a blend is tied to the goal of
the blending generation. A blend, or a part of it, may be more or less relevant
depending on what it is for. Once again, frames take a fundamental role as
being “context” specifiers, (i.e. the set of constraints within a frame describe
the context upon which the frame is accomplished). Therefore, having a set of
goal frames, which can be selected from the ones available in the Knowledge
Base or specified externally, a blend gets the maximum Relevance value if it
is able to satisfy all of them.
An aspect of the goal frames is that they become queries. For example,
if we want to find a concept that “flies”, we could build a goal frame with
the relation ability(x, f ly). The blends that satisfy this frame would be highly
relevant.

Definition 6.10 (Relevance). Assuming a set of goal frames, Fg , the set Fb of

the satisfied frames of blend b and the value PCNF for the pattern completion
of a set of frames F in blend b, we have
#(Fg ∩ Fb ) + #Fu × PCNFu
Relevance =
#Fg

where Fu , the set of unsatisfied goal frames, consists of Fu = Fg − Fb . This

formula gives the ratio of satisfied and partially satisfied goal frames w.r.t. the
entire set, Fg of goal frames.
From the point of view of creativity, we propose the use of Relevance as
a “usefulness” measure, an idea that will be applied in some experiments.

7. Elaboration

The Elaboration module is responsible for the application of several methods

of elaboration and completion to the blend. It is where the rules and frame
 Elaboration 137

conclusions are triggered and where the completion of uncompleted frames

(or patterns) is done. Hence, it encodes the elaboration function, θ , of the
model presented in the previous chapter. It allows the three different kinds of
elaboration explained: rule-based, internal logic and cross-concept based.
The rule-based elaboration, θR , is based on the application of rules from
the generic domain. Whenever the premises of any of these rules are proven
true, then their conclusions are inspected and the corresponding effects are
processed to the blend. As an example, we have the rule below, where we say
that, if an x lives in a house, and its habitat is water, then it must live in a
water tank, placed in the house:
rule(generic, water tank, [lives in(X, water tank), habitat(X, water)],
[lives in(X, house)],
[lives in(X, water tank),
in(water tank, house)],
[lives in(X, house)]).

If the premises are found to be true, then the Elaboration module will add
the relations lives in(X , water tank) and in(water tank, house) to the con-
cept map of the blend and also delete the relation live in(X , house). Another,
more complex example, is a rule that applies the movement laws to determine
the meeting point of two objects X 1 and X 2 moving on the same line:
rule(generic, meeting time, [starting position(X1, P0X1),
starting position(X2, P0X2),
{X1\=X2}, speed(X1, SX1), speed(X2, SX2), day(X1, D),
day(X2, D), starting time(X1, T0), starting time(X2, T0)],
[],
[{Dif is SX2-SX1, Dif\=0,T is (P0X1-P0X2)/Dif},
meet(X1, X2), time(X1, T), time(X2, T)],
[]).

The internal logic elaboration, θC , in this module inspects the frames that
are accomplished by the blend and in turn applies their effects, as with the
rules. In the special case of rules that are part of the blend (as a result of
projection from one of the inputs), we can also consider internal-logic elab-
oration, although the procedure that controls their application is the same as
for any other rules.
The cross-concept based elaboration is probably the least explored elabo-
ration method used. It is based on pattern completion. When the completion
evidence of a frame (as calculated in the Pattern Completion measure) sur-
passes a minimum specified value (the completion threshold), then the Elab-
oration module will add the missing relations that can be fully defined. For
138 Divago

example, in figure 31 (left), if the completion threshold was below 66%, then
the Elaboration module would add the relation ability(horse|bird, f ly). In
this case, the relation is fully defined because both arguments (horse|bird and
f ly) are known. If, on the contrary, the instantiation of the frame’s premises
had yielded ability(horse|bird, ) or ability(horse|bird, Ø), then no new re-
lation would be created. We call this cross-concept elaboration because it is
based on the transfer of knowledge from an external concept (the uncom-
pleted frames) to the concept map of the blend. We are aware that this may be
both an unsafe and incorrect way of doing cross-concept based elaboration. It
can be unsafe because, apart from the completion evidence, there is no other
method for ensuring correctness or meaningfulness of the added knowledge.
Only when dealing with goal frames (i.e. when there is an external motivation
to accomplish the frame), does the completion have a meaningful potential
consequence. It can be an incorrect perspective on cross-concept based elab-
oration because the source concept (the frame) is created from an analysis
of internal logic (the Pattern Completion measure), rather than being another
different concept that, for some reason, appears to be a good source of knowl-
edge.

The rules and frames applied in this module may mutually influence each
other. For example, the new knowledge added by a frame may in turn trigger a
rule and so on. This means that the system is sensitive to order of application.
To reduce this effect, the Elaboration module applies the rules iteratively until
no change to the concept map is made, i.e. until it stabilizes around a set
of relations. The drawback of this approach is that it becomes sensitive to
cycles. For example, suppose the following list of rules, with r1 , r2 and r3
being relations, and a blend containing the single relation r1 .

r2 ∧ not r3 ←− r1

r3 ∧ not r1 ←− r2
r1 ∧ not r2 ←− r3
After running the first rule, the Elaboration module would start by adding r2 ,
then would trigger the second rule, which would remove r1 and trigger the
third rule, which would return to the initial state, indefinitely. We are aware
that this is far from an unknown problem in the area of Logic Programming,
and thus we believe that good solutions may have already been found. How-
ever, for this book, and for the current version of Divago, the rules must be
 Elaboration 139

coded with attention to avoiding cycles, since we have not investigated farther
in this subject.
This module is very useful when one is knowledgeable about the domain
for which Divago is inventing concepts. For example, for the creatures exper-
iment, designed for a game project, we added some specific rules and frames
relating to solutions to problems (e.g. adding a wooden leg for a creature
when there is one leg missing) and for calculating values (e.g. calculating
a new strength value when there are conflicts between two possible candi-
dates).
The Elaboration module could have been applied after the GA cycle, to
the resulting best blend produced by the Factory, but this choice would imply
that the generation of the best blend itself could not take into account the
improvements from the Elaboration. In other words, the system would tend
to avoid generating solutions where, in spite of having originated from a low
valued non elaborated blend, the result after elaboration would compensate
its previous imperfections. This was the main reason for integrating it within
the GA of the Factory.
As with the Mapper, so the Elaboration module is optional for it can ob-
scure the inner workings of the blending mechanism in hiding imperfections
of the blend. Since this is important for the validation of the system, we also
allow the selection of the specific elaboration methods to apply.
Getting back to the model of concept invention that was the subject of the
previous chapter, we said that the elaboration function, θ , should be used to
define the space that would be traversed by the convergent strategy. Instead,
the Elaboration module is being (optionally) applied by the genetic algorithm
of the Factory, which we described as our divergent strategy. In this case, to
be coherent with ourselves, we ought to acknowledge that, when the Elabo-
ration module is used, we have a strategy that shares both the divergent and
convergent perspectives of the model. Thus, we present here no pure conver-
gent strategy. At first sight, this could be seen as a flaw in Divago and results
from the deliberate compromise of focusing divergence and bisociation and
disregarding other issues. However, and also for the same reasons, it became
gradually clear that the flaw is in the model itself. In fact, it seems much
more natural to consider convergence and divergence intermingled with each
other, rather than having a strict separation. We recall the analyses made in
section 1, in which we met the convergence/divergence dichotomy. Although
this duality was salient, in nowhere it was proposed to exist a strict separation
(rather, an interaction was often considered).
140 Divago

8. Divago as a creative system

To finalize this chapter, let us analyze Divago with the same criteria of section
2.2, where a classification of creative systems was proposed:
– Architecture. Single agent. As we have explained before, the approach that
we are following is centered on a single isolated system. In future stages, a
natural development would be to include it in a multi-agent environment.
– Model. Divago fits entirely the Cognition Centered Model (CCM)
paradigm, as it was developed from analyses of creativity from Cogni-
tive Science, Psychology, Philosophy and AI, and most of all it partially
integrates a computational implementation of a proposal for a cognitive
mechanism, named Conceptual Blending.
– Episodic Memory. Divago has NO true mechanism of Episodic Memory.
Although its implementation may possibly imply little more than a feed-
back loop (the outputs would become part of the KB), we are dedicated to
a feedforward version, as it showed sufficiently complex by itself.
– Evaluation. Divago has a built-in evaluation made by the Constraints mod-
ule. The only active participation of an external entity happens in the be-
ginning of the process (by setting the goal and the weights to associate
to each optimality constraint, and possibly also the input concepts and a
mapping).
– Theme. The theme of this project is Concept Invention.
– Reasoning paradigm. Divago is clearly a Hybrid system, in that it makes
use of rule-based reasoning, genetic algorithms and, to a much lesser de-
gree, connectionism (in the Mapper module).
– Domain. As we will see in the next chapter, Divago has been applied to a
variety of domains, namely 2D drawings, 3D creature design and linguistic
creativity.
Chapter 6
Experiments

Ever since its first sketches, Divago has been subject to experiments in several
domains. At the risk of making experimentation itself divergent, we decided
on this in-breadth approach for two reasons: we have been arguing in this
book for the consideration of multi-domain environments in computational
models of creativity; with an in-depth approach, we would focus more gradu-
ally on specific domain issues than on Divago itself. Whatever the approach,
validating this system with respect to creativity is a goal that we have been
pursuing, without finding any definitive and uncontroversial solutions as yet.
Because this is a fragile issue, we must follow the most solid principles we
can find. First, we need to avoid building the input knowledge structures in-
volved ourselves, the only exceptions being the first two experiments, the
house-boat and the horse-bird. Second, we try to read as little as possible
from the results except when there is a well-defined interpretation mecha-
nism. In other words, we try to avoid putting our own point of view on to
ambiguous events. Third, we seek to provide the statistically most signif-
icant data as possible to support the claims and conclusions achieved. We
follow the Central Limit Theorem, which says that “the distribution of means
from repeated samples from a population will be approximately normally
distributed if sample size is large (> 30) and exactly normally distributed if
the population itself is normally distributed”. Therefore, without knowing in
advance the distribution of the populations involved, we will rely on the con-
dition that each sample must be large. For example, each of the optimality
constraint weight configurations tested was subject to 30 runs with the same
starting conditions. Fourth, we estimate some of the measures for creativity
assessment, as presented in section 2, which will allow us to follow the same
evaluation framework throughout the experiments, which will be useful for
comparing the behavior of Divago within its own domains, as well as provid-
ing benchmarks for future comparisons with other systems.
The experiments are presented in chronological order and so the reader
will also perceive the evolution of the system in terms of the modules used,
the control over the results, the methodology and the interpretation of the
results. In the boat-house, we generate the whole set of possible combinations
(of boat|house drawings), using only the Mapper and the Blender. Only in
142 Experiments

the horse-bird experiments were we able to apply an objective analysis of

the results, when we applied the Factory and the Constraints modules for the
first time. There, we made the definition of novelty (the nov function) that
was applied to the rest of the experiments. The noun-noun experiments were
meant to test Divago with a large Knowledge Base and to compare it with
C3 , a Concept Combination system (see chapter 3). The creature generation
experiments are a study for the application of Divago to game environments
and also the first experiments with the use of the Elaboration module. Finally,
we apply Divago to some established blending examples in order to validate
it as a Blending model.
It will be obvious that, as we experiment with the system, some Optimality
Constraints are preferred over others, leading eventually to the elimination
of some and to the conclusion that one could reduce the list to a subset of
fundamental Optimality Constraints.
We will not hide the difficulties in analyzing Divago, namely in respect to
the value (or quality or usefulness) of the results, the evaluation of the fine-
tuning of the system or the individual effect of each of its components in the
results. Nevertheless, we hope to provide a set of objective conclusions and
benchmarks that may be useful for future comparisons.

1. The Boat-House

The first extensive experiment we made with Divago, published in (Pereira

and Cardoso 2002), had the goal of generating and analyzing the entire search
space from two input concepts. The context then given would be that of a sys-
tem with a specific goal (e.g. draw a “house”), but with a limited set of possi-
bilities (e.g. only one drawing example available), that would ask Divago to
extend its knowledge base of possible drawings.
In this experiment, the knowledge representation was restricted to concept
maps (built using Clouds (Pereira and Cardoso 2000)) and instances. Such
a limited representation was chosen because, apart from practical reasons
(this was the first experiment with Divago), an important decision was to
follow whose view on concept representation would be the center of further
developments (either micro-theory or exemplar view).
The choice for a “house” and a “boat” was made after some blending
and concept combination works (e.g. (Goguen 1999; Andersen 1996). We
intended to blend these two concepts and interpret the newly generated in-
 The Boat-House 143

Table 10. The house domain concept map

isa(house,physical structure) isa(human,mammals)
isa(physical structure, physical entity) isa(night, time object)
isa(time object, information entity) isa(skyscraper, physical structure)
isa(door, physical object) isa(window, physical object)
isa(roof, physical object) isa(observation, task)
purpose(roof, protection) isa(protection, task)
isa(body, physical object) isa(container, physical object)
isa(room, house part) isa(house part, space location)
isa(day, time object) isa(water proof, property)
isa(tree, vegetable) isa(vegetable, living entity)
live in(human, house) color(night, black)
have(house, door) have(house, window)
have(house, roof) have(house, body)
purpose(body, container) purpose(window, observation)
purpose(door, entrance) property(skyscraper, very big)
purpose(body, container) have many(skyscraper, house)
have many(house, room)

stances according to an unambiguous process. In this case, we decided to

define them according to a simple language (very similar to Logo (Abelsson
and diSessa 1981)), which enabled us to draw simple objects (a house and a
boat) and see the generated space without heavy computational work. In this
language, examples of commands are on/5, meaning “draw line for 5 pixels”,
off/5 meaning “move 5 pixels without drawing” or left/45, meaning “turn left
45 degrees”. This language as well as the syntax of the instances are described
in Appendix D.
The tables 10 and 11 show the concept maps of “house” and “boat”. A
short interpretation of these concept maps tells us facts like “a boat has a sail,
a hatch, a mast and a vessel, the vessel is the floating structure that serves
as container” or “humans live in houses, that have many rooms, a roof, a
window, a door and a body45 ”.
The Mapper generated 4 different mappings. In table 12, we show the two
mappings which most commonly appeared. While some concept mappings
come naturally (like “window-hatch” or “body-vessel”, in mapping 1), oth-
ers, less intuitively acceptable, appear as a consequence of the exhaustiveness
of the mapping function. For example, “water proof-slow” appears because
144 Experiments

Table 11. The boat domain concept map

isa(boat, physical structure) isa(sailing boat, boat)
isa(sail, physical object) isa(movement, task)
isa(triangle, geometric form) isa(geometric form, information entity)
isa(water proof, property) isa(hatch, physical object)
isa(observation,task) isa(mast, physical object)
isa(vessel, physical object) shape(sail, triangle)
shape(hatch, circle) have(sailing boat, sail)
have(sailing boat, hatch) have(sailing boat, mast)
have(sailing boat, vessel) have(vessel, floating structure)
purpose(sail, movement) purpose(hatch, observation)
purpose(mast, support) purpose(vessel, container)
property(sailing boat, slow) property(hatch, tiny)
property(boat,water proof) place(sailing boat, sea)
use(human, sailing boat) sail(human, sailing boat)

Figure 33. The boat and the house, as drawn from the instances

both can be “properties” of something (e.g. “physical structure can be wa-

ter proof”, and “boat can be slow”).
In table 13, we can see an excerpt of the blendoid corresponding to
mapping 1. According to it, possible relations in a blend could be that “a
house|sailing boat has a window|hatch that serves for observation, a door|sail
that serves for entrance|movement and has the shape of a triangle”, etc. No-
tice the combinatorial explosion that results from the choices given by the
blending projection operation.
Apart from definitions of shape that emerge in the blend (like,
shape(door|sail, triangle)), we don’t know exactly how to produce the vi-
sual re-interpretation of objects (e.g., what is the visual shape of door|sail?).
In other words, how can we read a blend? Since there is a (visual) precise
semantics for some of the concepts involved (such as roo f , sail or door), in
 The Boat-House 145

Table 12. Two mappings for the house-boat experiment

entrance ↔ movement body ↔ sail
task ↔ task container ↔ movement
protection ↔ support door ↔ hatch
roof ↔ mast entrance ↔ observation
door ↔ sail house ↔ sailing boat
house ↔ sailing boat physical
physical structure ↔ boat
structure ↔ boat window ↔ mast
window ↔ hatch roof ↔ vessel
body ↔ vessel water proof ↔ slow
water proof ↔ slow protection ↔ container
container ↔ container observation ↔ support
observation ↔ observation
1 2

Table 13. The blendoid concept map for house and boat
isa(entrance|movement, task) purpose(mast, support)
purpose(door, entrance|movement) isa(entrance, task)
purpose(sail, entrance|movement) isa(movement, task)
isa(roof|mast, physical object) purpose(door, entrance)
isa(roof, physical object) purpose(sail, entrance)
isa(mast, physical object) purpose(door, movement)
purpose(roof|mast, protection|support) purpose(sail, movement)
purpose(roof, protection|support) shape(door|sail, triangle)
purpose(mast, protection|support) shape(door, triangle)
purpose(mast, protection) shape(sail, triangle)
purpose(door|sail, entrance|movement) purpose(roof, protection)
have(house|sailing boat, body|vessel) have(house, body|vessel)
have(sailing boat, body|vessel) purpose(roof, support)
have many(house|sailing boat, room) have(house, body)
live in(human, house|sailing boat) have(sailing boat, body)
live in(human, house) have(house, vessel)
live in(human, sailing boat) have(sailing boat, vessel)
... ...
146 Experiments

Figure 34. Images that result from mapping 1

the form of Logo procedures, we must decide how to use them to produce the
new drawings. In the case where these concepts are found alone, their inter-
pretation is straightforward (just read the corresponding Logo procedures in
the potentially new context), but in the case of compounds (e.g. door|sail),
the problem becomes difficult. The ideal solution would be to find a way of
getting one degree of abstraction down and also blend the Logo procedures
themselves. However, the explorations done in this direction were leading
to a degree of complexity unjustifiable for the goals of the experiment. For
this reason, the interpreter made for these drawings ignored the compounds
and produced both alternatives (e.g. for door|sail, a drawing with door and
another with sail), thus producing the same results as with separate projec-
tions. Perhaps the most important conclusion from this experiment was that
the interpretation of a compound in itself opens up another blending problem,
recursively until a final and definitive answer is found (a possible solution in
drawings could be to apply visual morphing). In spite of reducing the search
space considerably by not proposing a different interpretation for compounds
and only taking into account the mappings that would have visual effects, a
large set of new drawings was produced. The mappings presented in table
12 generate respectively a set of 240 and 408 drawings with repetitions (giv-
ing approximately 80 and 100 different images, resp.). From mapping 1, we
found drawings such as those shown in figure 34.
By analyzing these house|boats, we can see some subtle transfers (e.g.
the square hatch in the first sail boat; the circular window in the house) and
 The Boat-House 147

Figure 35. Images from the mapping 2

Figure 36. Images from the body ↔ vessel, door ↔ mast and window ↔ sail blend

some blends that clearly share knowledge from both inputs, either visually
fortunate (e.g. the boat with rectangular sail) or unfortunate (e.g. the house
with the triangular door and a mast on top). It is also of relevance to say that
these unfortunate instances appear as a consequence of not having specific
domain-knowledge for generating a drawing or just because of unfortunate
combinations (second, third and fifth images).
When applying mapping 2, the results are as shown in figure 35. Notice for
example the different placement of the circle (and door). With the mapping 3
(body ↔ vessel, door ↔ mast and window ↔ sail), Divago produced draw-
ings such as in figure 36. Finally, the fourth mapping, which has a different
variation (body ↔ vessel, door ↔ hatch, roo f ↔ mast and window ↔ sail),
gave rise to images such as in figure 37.
148 Experiments

Figure 37. Images from the body ↔ vessel, door ↔ hatch, roo f ↔ mast and
window ↔ sail blend

The visual quality of the results would vary considerably depending on

the application of domain specific knowledge, such as guiding the result to
what a house or a boat should look like or which physical/structural rules they
should fulfill. Our goal with this experiment was to assess the generativity of
the system, regardless of any aesthetical judgment. Let us return to the context
given in the beginning of this section, a system with the goal of drawing
a house, and imagine the situation in which this system searches for house
drawings in the house domain, but cannot find any satisfactory solution. It
can then try to diverge gradually from the original domain (where novelty is
minimum), and get into a space of blends, where novelty increases, a sort of
middle space where concepts do not belong to a specific domain, but share
knowledge from more than one. In figure 38, we present this idea graphically.

Since we cannot have any precise measure of novelty or usefulness for the
drawings (that is not obscured by the subjectivity of the image) we cannot
do a thorough analysis of these results with regard to the criteria presented
in chapter 2. However, we can say that Divago produces a large proportion
(over 90%) of drawings that are definitely different from typical drawings of
houses or boats (they gather different elements of house and boat in the same
drawing, sometimes lacking some parts or violating basic drawing principles,
like non superimposing objects) in a total of more than 1300 drawings (with
repetitions), if counting all the mappings. This basically corroborates that, far
from reinventing (i.e. converging), instead it generates novel, yet potentially
pointless, results (i.e. diverging).
In general, we can say that, although the two concepts of a house and a
boat are close to each other (both are physical structures, used by humans),
this can be an example of computational modelling of divergent thought be-
cause a large amount of new instances was generated from the blending of
 The Horse-Bird 149

ZOOM

Domain B

Domain D
Space of Blends

Domain A

Domain E
path towards the solution
starting point of search
solution found

Figure 38. A search in a multi-domain environment

two different concepts. According to this perspective, Divago could serve as a

meta-level engine for helping another system with extending its search space.
This is particularly feasible in situations where the search space consists of a
set of independent knowledge structures, such as in Case-Based Reasoning.

2. The Horse-Bird

The Horse-Bird experiment was the first to assess the behavior of the Factory
and Constraints module (presented in (Pereira and Cardoso, 2003a)46 ). Ac-
tually, it consists of two different kinds of experiments, each with a distinct
goal: assessment of the individual effects of each measure on the final results;
qualitative evaluation and tuning of the model. After several preliminary GA
parameters tuning tests, we decided for 100 individuals as the population size,
5% of asexual reproduction (copy of an individual to the following popula-
150 Experiments

tion), 80% of crossover (combination of pairs of individuals), 14% of mu-

tation and 1% of random generation (to allow random jumps in the search
space)47 . We have three different stopping conditions: appearance of an in-
dividual with the maximum value (1); achieving n populations (n = 500);
being stalled (no improvements in best value) for more than m populations
(m = 20). We kept these GA configurations throughout the whole experimen-
tation related to the Horse-Bird blending, as described here.
As a result of the problems relating to the interpretation of compounds
in the house-boat experiment (e.g. window|hatch), we decided to drop this
kind of projection, thus reducing the search space, now having from 2l to
32k × 2l−2k different blends. Notice that the concept maps of horse and bird
(already given in tables 6 and 7) have (m=) 29 and (n=) 33 different concepts,
respectively, which gives an l=m+n=62. The three mappings used (already
given in figure 9) have sizes k=6, k=5 and k=21, thus the search space will
have a minimum size of 262 and a maximum size of 342 × 220 . Even if dis-
counting the concepts and relations that are repeated in both concept maps
(e.g human), this corresponds to a very large set of blends.
This experiment also introduces the criteria for defining novelty (the nov
function, converse of Ritchie’s typ function) which will be used throughout
the rest of this chapter. This function is based on the comparison of the con-
cept map of the blend with those of the inputs. The exact value of distance
between an input x and the blend b corresponds to the sum of the relations
that belong to b and that are missing in x with those that belong to x and
are missing in b. Let us call it d(b, x). This can be seen as an edit distance -
the set of delete and insert operations needed to transform one into the other.
Since this value becomes proportional to the sizes of the concept maps in-
volved, we divide it by the size of the blend concept map (the number of
relations), thus getting a normalization that allows us to compare among dif-
ferent experiments and assess the behavior of the system. Following one of
the measures of comparison to an archetype (novelty2(x)) by (Pease, Win-
terstein and Colton 2001), we define the function distance as returning the
normalized minimum distance to one of the inputs:

min(d(b, x1 ), d(b, x2 ))
distance(b) =
sizeb

such that x1 and x2 are the input concepts for generating b and sizeb is the size
of the concept map of b. The larger the distance to inputs, the higher is the
 The Horse-Bird 151

novelty, therefore, the function nov is defined as:


1 distance(b) > 1
nov(b) =
distance(b) otherwise

Determining usefulness (use) presupposes the existence of a purpose.

Therefore, we will only apply it when this is explicit in the experiment. In
the second part of these experiments, we have the goal of finding a pegasus,
and the distance to this point in space will give us an estimate to use. When
analyzing the set of values for nov and use given in a sequence of runs, we
tend to prefer the median since it is not sensitive to outliers, as happens with
the mean, and it normally represents a specific blend that is representative of
the mean and that we can inspect. Whenever this assumption becomes un-
safe, if there is a large difference between the median and the mean and a
large standard deviation, we will also consider other indicators. In any case,
the reader will find the values for median, mean, standard deviation and mode
in the result files (found in appendix E). Therefore, unless stated otherwise,
the statistics presented refer to the median of the results for 30 runs. The in-
spiring set (following section 1) will comprise the two creatures, horse and
bird, as well as the pegasus, since the latter was also explicitly defined by us.

2.1. Evaluating the Optimality Pressures

This experiment serves to observe the effect of each pressure in the final
results, bringing up a way to predict and control the system. For the first part
of these experiments, we isolated each optimality pressure, by attributing zero
weight to the remaining criteria. Since one of the optimality pressures is not
independent (Web) and another (Intensification of V.R.) was not accounted
for in our current implementation we have only six different criteria to take
into account.
The input domains applied were the domains of horse and bird (in ta-
bles 6 and 7), meaning that the expected results range from the unchanged
copy of one (or both) of the concepts to a horse-bird (or bird-horse) which is
a combination of selected features from the input domains. The generic do-
main consists of the general ontology, integrity constraints and a set of frames
(already given in table 8; see also Appendix E).
We applied the three mappings presented in figure 9. For each mapping,
we tested the six optimality pressures, each of these comprising 30 runs48 .
152 Experiments

The Elaboration module was not used. Each blend was examined by the Con-
straints module without being subject to any transformation after the projec-
tions.
We now present an analysis of the individual effect of each of the mea-
sures:
– In Integration, frames behave as attractor points in the search space. More-
over, the frames with a larger coverage tend to be preferred, although when
too large (like aprojection or aframe) they are dropped. The evolution is di-
rected to a compromise of coverage and satisfaction. The complexity of the
search space grows with mapping size (the number of cross-space associ-
ations found by the mapping algorithm). In fact, when we have a mapping
of size 5, six different blends are returned, the best choice being retrieved
43% of the times, while with a mapping size of 21, eight different solu-
tions are found, the best choice being retrieved a mere 6% of the time. This
confirms the complexity and dimensions of the search space we discussed
in section 5. A good compensation for this apparent loss of control is that
the returned values are clearly higher (0.68, for the best) than in the small
mappings (0.22), suggesting that, with larger mappings, the probability of
finding a better solution is higher than in smaller ones. Finally, the novelty
was 0.71, meaning that the set of frames used does not lead naturally to
any of the inputs, i.e. the system diverges from its input concepts.
– Pattern Completion drives the blend to partially complete (i.e. satisfies
some conditions but not all) the highest possible number of frames, lead-
ing, in each case, to several sets of relations which fit into those frames
without satisfying them completely. This means that, isolated, Pattern
Completion only leads to disperse, non-integrated results and so it is not
very useful. Interestingly, it can be useful when combined with Integration
because it gradually brings to the blend the concepts and relations that are
needed to complete the frames and so it speeds up the process of finding
frames with high Integration value. In respect to the search landscape, it
seems to be very rich in local maxima. The most constant results came
from mapping 2 (of Figure 9), with the best results obtained 13% of the
time and the second best 20% of the time. An interesting remark is that
the local maxima always fall within a very strict range of values (of max-
imum amplitude 0.11, in mapping 3). The median value for nov was 0.79,
which confirms our expectancy that Pattern Completion would be close to
Integration, in these terms, since they use the same set of frames.
– From the experiments with Topology, we can observe that there is a ten-
 The Horse-Bird 153

dency to bring all the relations from both concept maps to the blend, with-
out being transformed. This means that, at the limit, the blend will com-
prise the union of the two concept maps from the inputs, thus (if both the
inputs have the same size) the novelty will tend to be 0.50 (half the concept
map of the blend would have to be deleted to become an exact copy of one
of the inputs). This prediction is corroborated by the result (nov= 0.51).
– The influence of Maximization of Vital Relations in the results is straight-
forward, given that its highest value (1) reflects the presence, in the blend,
of all the vital relations that exist in the inputs. As the evolution goes on
in each run, the value grows until reaching the maximum reasonably early.
For each set of the 30 runs, it reached the value 1 a minimum of 93%
of the times, and the remaining 7% achieved at least a value of 0.95. As
in Topology, the search space of Maximization of Vital Relations is very
simple since there is a global maximum in the neighborhood of (almost)
every point. However, in contrast to Topology, this measure results in very
high novelty (0.99), which can be explained by the fact that the number of
vital relations in the concept maps is relatively small and that there is no
constraint on the arguments of these relations. In other words, it does not
matter what the vital relations actually associate with, only that their sim-
ple presence in the blend is important in order to get the maximum value in
this measure, yielding an apparently random choice of elements projected.
– The results of the Unpacking measure show that it drives towards similar
results as Topology, with the main difference being that the relations in the
blend are clusters of copies of subgraphs from the inputs. I.e., Unpacking
only copies those relations that do not imply conflicts (e.g. some concepts
that belong to both domains, such as leg, can become problematic because
its Unpacking is unambiguous). It is therefore a force of inertia. The me-
dian value for nov was 0.63, testifying that, whatever was present in the
concept map of the blend, it was similar to one of the inputs yet missing
some parts that would make it an exact copy, which intuitively agrees with
its definition.
– The first part of the test on Relevance focussed on making a single relation
query. In this case, we asked for “something that flies” (ability( , f ly)). The
results were straightforward in any mapping, accomplishing the maximum
value (1) in 100% of the runs, although the resulting concept maps did not
reveal necessarily any overall constant structure or unity, giving an idea of
randomness in the choice of relations other than ability( , f ly). In other
words, the evolution took only two steps: when no individual has a rela-
154 Experiments

tion ability( , f ly), therefore with value 0; when a relation ability( , f ly)
is found, yielding a value 1, independently of the rest of the concept
map. The second part of the test on Relevance, by adding a frame (abil-
ity explanation) to the query, revealed similar conclusions. There was no
sufficient knowledge in any of the input domains to satisfy this new frame
completely, so the algorithm searched for the maximum satisfaction and
reached it 100% of the time in every mapping. So the landscape seems to
have one single global and no local maxima, reflecting the integration of
the two parts of the query. The existence of local maxima would be ex-
pected if there were separate frames. Intuitively, the search landscapes of
Integration and Relevance seem to be similar. As with Integration, the nov-
elty is dependent on the available frames, more specifically on the frames
used in the query. With the ones used, the value for nov was 1. This is con-
sistent with the observations just made of the apparent randomness of the
choice of relations for complementing the concept map.

2.2. Finding the Pegasus

For our concerns, we define a pegasus as being a “flying horse with wings”, so
leaving out other features it may have (such as being white). These extra fea-
tures could also be considered but would need knowledge about any aspects
of ancient Greece, Greek mythology and some ontological associations (e.g.
purity is white). Moreover, they would make the generation of the blend con-
siderably more complex, although possibly more interesting. Formally, the
pegasus we want to generate has the same concept map as the horse domain
augmented with 2 wings and the ability to fly (so, it should also have the
relations ability(horse, f ly), motion process(horse, f ly), pw(wing, horse),
quantity(wing, 2) and purpose(wing, f ly)).
For validation purposes, we started by submitting a query with all the rela-
tions of the pegasus, so as to check if they could be found in the search space,
and the results reveal that only the mapping 3 (see figure 9) respects such
constraints. This led us to use this mapping exclusively throughout the rest of
the experiment. Knowing that the solution exists in the search space, our goal
was to find the minimal necessary requirements (the weights, the frames and
the query) in order to retrieve it. From a first set of runs, in which the system
considers a large set of different frames and no query, we quickly understood
that it is not simple (or even feasible) to build the pegasus solely by handling
 The Horse-Bird 155

the weights. This happens because the optimality pressures provide control
regarding to structural evaluation and general consistency, but only by pure
chance can we find the exact weights to match the same relations of the pega-
sus, a very specific blend that fails to follow all but a few of constraints, but a
combination of them. This drives us to the need of queries.
A query may range from specific conditions that we demand the blend
to respect (e.g. the set of conditions for flying, enumerated above) to highly
abstract frames that reflect our preferences in the blend construction (e.g. the
frame aprojection: elements from input concept map 1 should all be pro-
jected). Intuitively, the best options seem to comprise a combination of the
different levels of abstraction.
Since a query is only considered in the Relevance measure, its weight
must be large if we intend to give it priority. In fact, using only Relevance
is sufficient to bring the concept map of the solution to the blend, when the
query is specific enough, as we could test by using a query with aprojection
and the flying conditions. From a creativity point of view, it is not expected
to have very specific queries and we are more interested in less constrained
search directives. In table 14, we show the parameters we used, as well as the
nov and use values obtained. use is calculated as:
d(b,target)
use(b) = 1 −
sizeb
with target being the concept map of an optimal blend (in this case, the Pega-
sus). The weights we present correspond to Integration (I), Pattern Comple-
tion (PC), Topology (T), Maximization of Vital Relations (MVR), Unpacking
(U) and Relevance (R). The fly conds. are the relations that the blend must
have in order to be a flying creature, and aframe, aprojection and new ability
are frames as described before (and detailed in appendix E). The values pre-
sented correspond to the median in each set of results.
An observation that must be made is that the target is very similar to one
of the inputs (the “horse”), its novelty being exactly of 0.26, a very low value
that was only acknowledged after the first experiments. Since making it less
typical would imply artificial changes in the concept map (actually the Pega-
sus is a horse with wings), we decided to leave it untouched. Furthermore, it
is theoretically possible to generate a blend that is close to the pegasus, yet
far away from the horse (if it falls in the opposite direction of similarity). As
we can see from the experiments, there are useful results that nevertheless fail
the threshold of novelty and there is no linear relationship between nov and
156 Experiments

Table 14. The 10 different configurations used. Query: A=fly conds. + apro-
jection; B=fly conds. + aframe; C=fly conds.+ aprojection + aframe;
D=new ability+aprojection+aframe
Exp. Weights Query nov use Best blend
# I PC T MVR U R (nov/ use)
1 0 0 0 0 0 100 A 0.59 0.53 0.40/0.74
2 0 0 0 0 0 100 B 0.86 0.26 0.0/49
3 0 0 0 0 0 100 C 0.59 0.53 0.40/0.71
4 50 0 0 0 0 50 C 0.51 0.62 0.19/0.97
5 33.3 33.3 0 0 0 33.3 C 0.78 0.34 0.82/0.32
6 33.3 0 33.3 0 0 33.3 C 0.60 0.52 0.49/0.66
7 25 0 25 25 0 25 C 0.70 0.28 0.45/0.58
8 20 0 20 20 20 20 C 0.62 0.33 0.47/0.51
9 34 0 16 10 4 36 C 0.43 0.70 0.44/0.95
10 34 0 16 10 4 36 D 0.26 0.71 0.35/0.73
11 20 0 0 0 0 80 C 0.16 0.76 0.21/0.90
12 20 0 0 0 0 80 D 0.58 0.47 0.18/0.92

use, although when use gets high scores, the opposite happens with nov.
The first eight configurations were dedicated to understanding the effect
of gradually adding optimality pressures to the fitness function. In the first
three, where only Relevance was used, we verified that, although it was easy
to have all the concepts and relations we expect for a pegasus, often it was
complemented by an apparently random selection of other relations. This re-
sults from having no weight on Integration, which we added on the config-
uration 4, yielding the result that was closest to our pegasus: the projection
of the entire horse domain, and the selective projection of wings and the f ly
ability from the bird domain. There were a few extra bits of knowledge, such
as having two claws, feathers or chirping. The majority of the time, the extra
knowledge results in blends that are distant to the inputs and to the pega-
sus, i.e. the pegasus found was more a singularity than the average situation.
The straight explanation is that the weight of Integration leads Divago to sat-
isfy frames that compete with the pegasus (e.g. bframe, which would project
the bird’s concept map structure) in many different ways. In configuration
5, the influence of Pattern Completion led the results to minimum incom-
pleteness (e.g. a pegasus with everything except a mane, wings or any other
item), which revealed that, by itself, it is not a significant or even positive
contribution to the present goal, a reason for dropping its participation in the
subsequent configurations.
Adding Topology (conf. 6) essentially brought two different kinds of re-
sults. As with configuration 4, it returned the “correct” pegasus with extra fea-
 The Horse-Bird 157

tures like having f eathers or a beak, each of which was apparently selected
at random. These were also given the highest scores in the fitness function.
However, in some of the runs (10%), the results contained both creatures
(horse and bird) in the same concept map, as if they were connected (e.g.
having the same legs or ears). This is a rather unwanted result, and it sug-
gests that the weight of Topology should be relatively small in comparison to
others.
The following configuration, the inclusion of Maximization of Vital Re-
lations, confirmed the same conclusions as from Topology, but with more
control over the kind of extra relations transferred to the blend. For example,
the blend may have 2 wings (from the relation quantity), a beak and feath-
ers (from pw), but it is never an oviparous (from member o f ). On the other
hand, we can sense a gradual lack of focus on the overall results (no two runs
returned the exact same result) complicating considerably our goal of control-
ling the system. There is a simple explanation for this: Relevance, Integration,
Topology and Maximization of V.R. all have the same weight and some (like
Maximization) are more easily satisfied, thus driving the evolution towards
their maxima, from wherever the evolution started. This same phenomenon
happened in configuration 8, although Unpacking had brought a more stable
set of results.
An immediate conclusion we took from these experiments was that each
pressure should have a different weight, correspondent to the degree of influ-
ence it should have in the result. In our case, we are seeking for a specific
object (the pegasus), we know what it is like, what it should not have and
some features not covered by the query conditions that we would like it to
have. This led us to a series of tests for obtaining a satisfiable set of weights,
used in the configurations 9 and 10. Given the huge dimension of the problem
of finding these weights, they were obtained from a generate-and-test process,
driven by our intuition, so there is no detailed explanation for the exact choice
of why these values and not others. Yet, a qualitative analysis can be made
and we see a clear strength given to Relevance and Integration. The former
serves to “satisfy what we asked” and the latter guarantees overall coherence
(centered on the query frames) and consistency (e.g. it prevents the solution
from having 2 and 4 legs simultaneously). There is also a more discrete pres-
ence of Topology, Maximization and Unpacking, to allow the transfer of extra
knowledge. Configuration 9 revealed, possibly, the “richest” pegasus found,
in the sense that, although largely failing the target, it contains all of its rela-
tions as well as a selection of other relations (having lungs, f eathers, a pair
158 Experiments

of claws). Still, although this result appeared consistently throughout some

runs, there is a high variability of results (for configuration 9, the mean of
nov was 0.64 with standard deviation σ = 0.17; for conf. 10, the mean for
nov was 0.57, σ = 0.17) testifying the difficulties in controlling the system.
Finally, from what we had learned in the first configurations, we decided
to reduce to only two constraints (Relevance - 80% and Integration - 20%),
predicting that we would find the best approximation to the target. Indeed,
this was confirmed by the results, particularly for configuration 10. Although
still not avoiding very bad outliers in two runs which seriously affected
some indicators (it yielded meannov = 0.24, σnov = 0.08 and meanuse = 0.70,
σuse = 0.30), the results were very stable (if removing the 6% outliers, we get
meannov = 0.16, σnov = 0.06 and meanuse = 0.77, σuse = 0.10). We will use
this specific configuration later on when we know with some confidence the
kind of output to get (and we have the necessary frames).
Now, for a more detailed analysis, we will calculate the values for
Ritchie’s measures (or criteria). As we have referred to before, we assume a
mapping between the pairs novelty/usefulness and typicality/value, such that
novelty is the opposite of typicality (typ=1−nov) and usefulness equals value
(val=use). The latter may become controversial, yet it may be the best method
for applying Ritchie’s measures in this context and, above all, it is our convic-
tion that, for a formal setting such as the one we are describing, one can only
measure the value of something as much as it accomplishes a goal or satisfies
a set of conditions. In other words, it must be a solution to a problem, i.e.
be useful. According to this philosophy, we obtain the values in table 15. We
remember that we have assumed the value 0.5 for the parameters α , β and γ
From the first four measures, we can say that Divago is producing typical-
ity and value near the half scale (measures 1 and 3), but clearly produces more
valued than typical items (2 and 4). From measure 5, we can see that all typ-
ical results were valued, which is clearly due to our target falling within the
range of typicality. However (from 6 and 7) there are some valued but non-
typical outcomes, which is a good indication of creativity. The proportion of
valued non-typical outcomes with regard to the typical ones (measure 8) can
be misleading since we are comparing results with different configurations
- the last 4 configurations clearly yield more typical, valued results than the
others. Reinventions only occurred in occasional runs, which do not fall into
the statistical validity, therefore we cannot say that Divago has consistently
produced reinventions, as can be seen by the measures 9 and 10. This fact is
important in understanding why measures 11 to 14 are reproductions of the
 The Horse-Bird 159

Table 15. Ritchie’s (Ritchie 2001) measures results

Measure Value
1 0.443
2 0.273
3 0.504
4 0.636
5 1.000
6 0.364
7 0.500
8 1.333
9 0.000
10 N/A
11 0.406
12 0.483
13 0.273
14 0.636

first four. Indeed, there is such a correlation between these eight measures
that we suspect they can be reduced to a smaller set.
For determining the fine-tuning of the system according to (Colton, Pease
and Ritchie 2001) (see Appendix A), we can at most determine an estimate,
given the participation of so many variables in the definition of each result.
From the experiments so far, we can say, for example, that Relevance and In-
tegration are creatively more useful than Pattern Completion, and that aframe
is less creatively useful than aprojection (compare configurations 1 and 2).
However, we can also see the high complexity involved. Compare, for exam-
ple, configurations 11 and 12. The former normally produced better results
than the latter, but the latter has a best blend with higher scores. It seems,
therefore, that configuration 11 earned more stability at the cost of losing
better singularities to the configuration 12. To check the individual influence
of the frames in configuration 11, we applied the same weight configuration
of 80% Relevance and 20% Integration to all possible combinations of the
query (see table 16). Notice that the first two queries (void query and fly
conds.) completely fail to achieve anything useful. This results from not hav-
ing applied an organizing frame (such as aframe or aprojection). In this case,
Divago was directed towards satisfying anything (empty list) or just a small
160 Experiments

Table 16. Checking the fine-tuning of Divago

Best Blend
Query nov use nov/use
empty list 1.00 0.00 1.00/0.00
fly conds. 1.00 0.00 1.00/0.00
aprojection 0.36 0.56 0.19/0.74
aprojection, fly conds. 0.35 0.81 0.21/0.95
aframe 0.00 0.74 0.00/0.74
aframe, aprojection 0.15 0.67 0.00/0.81
aframe, aprojection, fly conds. 0.16 0.76 0.21/0.90

set of relations (fly conds.), preventing it from using an organizing frame and
therefore making a coherent whole. From this, we can conclude that aframe
and aprojection are important to organize the blend (in this case, towards the
same organization of the horse concept map).
We can also notice that, when using only aframe and aprojection (aframe
alone or combined with aprojection), Divago produces either an exact copy
of “horse” or a very similar result. However, when put together with the flying
conditions, it makes a whole that can lead to (very nearly) the pegasus. This
may indicate a high fine-tuning towards the pegasus, however we can also see
that the same combination can lead to other results depending on the weights
applied (e.g. configuration 5) and other combinations (e.g. configuration 12)
can lead to the same results with the same weight configurations. To conclude,
if, on the one hand, the frames are a method for controlling/tuning the system,
it is also true that their application does not guarantee valuable results and
that, in this sense, fine-tuning the system is an extremely difficult task. This
results from the complexity of the space and from the specificities of frame
combinations (some may be compatible, some others may be competing).
It is clear that the results in this section were driven subjectively by us in
the choice of the concepts and frame design, but the argument we are trying
to make is that we can lead Divago to produce novel and useful outputs.
Nevertheless, it is a difficult system to control, a good aspect on one side -
it is hard to be intentionally biased to specific outputs -, but bad on the other
side - it is extremely difficult to test its full potential.
We also developed (in collaboration with Pablo Gervás (Pereira and
Gervás 2003)) an Interpreter for generating textual descriptions of the blends,
based on Natural Language Generation techniques. This system made de-
 Noun-Noun combinations 161

scriptions by comparison with the input concepts of “horse” and “bird”. Ex-
amples of automatically generated descriptions of blends are:

(1) A horsebird is a horse. A horsebird has two wings and feathers. It can fly,
and it moves by flying.

(2) A horsebird is a horse. A horsebird can fly, it has feathers, a beak, and
wings for flying and it moves by flying.

(3) A horsebird is a horse. A horsebird can fly. It chirps, it has wings for flying
and it moves by flying.

The example (1) corresponds to a result from configuration 4. Examples

(2) and (3) are interpretations from configuration 9.

3. Noun-Noun combinations

In the experiments with noun-noun combinations (published in (Pereira and

Cardoso 2006)), we show the behavior of Divago with a dataset constructed
independently by other researchers (Costello 1997) and make a comparison
to C3 (see section 3.1). Each concept (associated to a noun) is represented
with a syntax that is equivalent to the one adopted for Divago. Here, we ap-
ply for the first time the two-step methodology that will be followed in the
subsequent experiments, which starts by “tuning” the system with preferred
outcomes and then allowing it to do free generation, constrained by a given
query. Also for the first time, we define a more precise criterion for usefulness
(use), which will correspond to the score obtained for the Relevance princi-
ple. The rationale is that a blend is useful (i.e. valued) if it accomplishes a set
of pragmatic conditions that may be specific to a situation (e.g. it should be
a clay object that serves to cut food) or a generic demand for an application
(e.g. it should be an object with a single color and a single shape), which can
be given as a query to the system. Thus, use will now correspond to the value
of Relevance. This may seem contradictory with the choice for use in the pre-
vious experiment, but the only difference is that now use will participate in
the search, which goes in agreement with the model of creativity presented,
in which the invention of new concepts should be purpose-driven. The value
for nov will be given exactly as before.
The dataset used in these experiments comprises 179 concepts (noun con-
ceptual descriptions) borrowed from Fintan Costello’s PhD thesis (Costello
162 Experiments

1997) on noun-noun conceptual combination. In this thesis (and in subse-

quent publications (e.g. (Costello and Keane, 2000)), the author describes
each concept by a set of attribute-value pairs, as shown below (for “neck-
lace”)

Necklace
name: (necklace)
feature-set: (solid inanimate static)
color: (silver gold)
shape: (small circular)
structure:
made of: metal
parts: (pendant)
found:
function: ((wears person3 necklace neck)
(decorates necklace person3))

The conversion to our concept maps is straightforward: each of the “fea-

tures” becomes a property relation; the attributes color, shape and made o f
become relations with the respective name; each of the parts is converted into
a pw (part whole) relation; each “function” is converted into a set of actor and
actee relations (with third arguments, such as place or instrument). The actor
is expected to be the first argument of the function, while actee is the second.
Therefore, our concept map representation for “necklace” is as follows:

property(necklace, solid) made of(necklace, metal)

property(necklace,inanimate) pw(pendant, necklace)
property(necklace,static) actor(wears, person3)
color(necklace,silver) actee(wears, necklace)
color(necklace,gold) place(wears, neck)
shape(necklace,small) actor(decorates, necklace)
shape(necklace,circular) actee(decorates, person3)

In the original dataset, there are interrelationships between nouns. For

example, there is also a representation for pendant, person3 and neck, so,
 Noun-Noun combinations 163

along with necklace, these nouns can be seen as a small graph representing
the knowledge about people and necklaces. Within this small graph, there is
normally no repetition of function specifications (e.g. in neck or person3 rep-
resentation, there is no wears function, although it exists implicitly). For this
experiment, we directly and separately converted each noun to a concept map,
and there is no communication between our concept maps, which means that
many nouns in our knowledge base lose their original implicit data. This was
necessary since automatically converting that implicit network into our con-
cept maps would not be a trivial exercise in terms of programming and would
clearly fall away from the goals of this project. Another aspect of the dataset
is that some concepts have several different instantiations (e.g. person3 is the
third representation of the noun person). We also converted these directly and
separately to our knowledge base, without merging them.
The main goal of these experiments was to observe how Divago behaves
with respect to criteria of novelty and usefulness when applied to knowledge
from another concept combination system. Another intention was to improve
the control over Divago with regard to these measures.
The noun-noun interpretations we consider in the experiments are either
hybrid interpretations or property interpretations (see section 3.1). In some
tests we made prior to these ones, the Mapper (which is based on structure
alignment) was clearly unable to allow other types of interpretations such as
relational and known-concept interpretations. This may point in the same di-
rection as Costello and Keane (Keane and Costello 2001), who argue that con-
ceptual combination cannot be reduced to structure alignment. In our case,
the strictness of the structure alignment methodology affects Divago’s needed
flexibility: 1-to-many mappings should sometimes be considered (rather than
1-to-1); representation of inputs sometimes with variable granularity should
be possible (e.g. “fido is a canine, who is a mammal” mapped to “tweety
is a bird” should yield f ido ↔ tweety and mammal ↔ bird, rather than
canine ↔ bird or canine ↔ tweety). Still, we cannot argue that these lim-
itations are more than computational and/or representational limitation, so
further exploration regarding structure alignment and computation should be
taken before claiming it as too rigid.
In order to provide a pragmatic background for the experiments, we invite
the reader to consider a situation where one wants to obtain combinations
with a specific set of characteristics, we can thus define this set via scripts
with the same syntax of the nouns described above. A useful concept must
have specific values for the slots of the script and respect a set of integrity
164 Experiments

constraints. The slots and values required can thus be grouped together in a
query. In all experiments (except in the tuning set), this query consisted of:

property(A,[animate, inanimate]),
property(A,[liquid, solid]),
property(A,[static, mobile]),
made of(A, ),
shape(A, ), color(A, ),
actor(F, ), actee(F, )

Square brackets mean disjunction (e.g. the concept A must be animate or

inanimate). The presence of actor and actee relations means that the concept
should have a function.
Beforehand, we could not know exactly which kinds of frames were
needed to build “good” combinations, leading to the need of a tuning phase
that helped us find a set of appropriate frames. Only after this tuning, are we
able to test the system, leaving it to construct its own concepts.

3.1. Tuning

The tuning set we used consisted in 30 pairs of randomly selected concepts

from the list. For each one, we constructed a solution (called the tuning target)
correspondent to our own interpretation of the noun-noun combination. This
hybrid interpretation considered exclusively the knowledge contained within
the selected noun representations and was centered on the head noun, which
means that, in any pair A-B of nouns, the interpretation was that “an A-B is
a B with such and such A characteristics”. In other words, the concept B, the
head, is always the focal concept in our interpretations.
Each experiment consisted of making 30 runs for each pair (each run with
the exact same starting conditions), having in the query the set of frames that
could be expected to achieve the target. The weight configuration followed
was 90% for Relevance and 10% for Integration, which reflects our intention
to test the frames. When the results were mostly missing the target, we ei-
ther selected other frames or designed new ones and made the 30 runs again.
More specifically, this happened when there was an error of more than 2 re-
lations to the target or when this error was due to fundamental relations (i.e.
 Noun-Noun combinations 165

without them, the result would not be novel or valued). In table 17, we show
a sample with the tuning combinations, target descriptions, resulting differ-
ence to the target, novelty score and frames used in the query. It is important
to remember that the target interpretations are obtained using only the exist-
ing knowledge representation of both nouns, which justifies the appearance
of awkward interpretations (e.g. “head hammer handle”, “pen person”). We
can also see that the frames were initially tailored to fit the target interpreta-
tions and reused later when effective (e.g. the “shape transfer” was created
for “bullet potato”, and used often in the succeeding experiments).
The table 18 presents the frames that were obtained (or selected from the
already existing ones in the generic domain). For the rest of the experiment,
this became the set of available frames.
It is clear though that both the target interpretations and the frames were
made by us, so introducing a subjectiveness component in these experiments.
Since there does not seem to be any simple automatic frame generation mech-
anism and given that the language itself demands some expertise, the frames
had to be constructed with the method described. On the other hand, it would
be possible to use other people’s interpretations of the randomly generated
pairs, requiring a reasonably large set of participants with some expertise to
understand the constraints (interpretations are confined to the specific repre-
sentation). This was done for the next two experiments (established blends
and game creatures). Not having done so for this experiment, we tried to fol-
low our intuition and imagination in each case. At worst, the experiments
reflect our specific ways of noun combination on the tuning set applied to the
free generation set.
The mappings used in all the experiments were automatically generated
by our structure alignment algorithm, with the seed dormant bridge connect-
ing the individual identifier symbol of the nouns (for example, in “necklace
paper”, the seed dormant bridge is necklace and paper, which then goes to
the made o f relations, establishing a mapping between metal and paper and
so on). It typically established mappings between elements with the same role
in both nouns (color value with color value, made o f value with made o f
value, etc.)
166 Experiments

Table 17. Excerpt of the tuning set (average distance to target (average error)=0.67,
standard deviation=0.994)

Combination Tuning Target Error Frames nov

bullet potato small and cylindrical potato 2 bcore. shape transfer. 0.08
slot set completion
cow vehicle body black and white vehicle that eats grass 0 bcore, function transfer, 0.44
slot set completion
eagle shirt brown, bird-shaped shirt 1 bcore, structure transfer, 0.46
shape transfer
engine ball self-mobile ball 0 bcore, feature set contrast 0.29
flower bloom plant spherical plant 0 bframe, shape transfer 0.12
fruit1 paper1 paper with fruit-seeds that humans eat 0 bcore, structure transfer, 0.40
function transfer
head hammer mobile and animate (living) hammer 2 bframe, feature set 1.00
handle1 handle contrast
neck instrument small and straight instrument 0 bcore, shape transfer 0.25
necklace paper circular paper that people use in the 1 bcore, function transfer 0.45
neck for decoration
patient paper1 paper that has illness 0 bcore, function transfer 0.14
pen person thin, long person, that is used (by 2 bcore, function transfer 0.67
others) to write on paper
pencil pendant thin, long pendant, used to write 0 bcore, shape transfer, 0.46
on paper function transfer
potato acorn brown, spherical acorn 1 bframe, function trasnfer 0.30
potato herring tail spherical herring tail 2 bcore, shape transfer 0.12
pottery spoon spoon made of clay 1 analogy transfer 0.56
skin stem thin stem 0 bcore, shape transfer 0.14
spoon1 frame brown frame 1 bframe, single 0.00
differentiating feature
spoon1 handle lens straight and long lens 0 bcore, shape transfer 0.25
thorns hammer1 small and sharp hammer 0 bcore, shape transfer 0.13
tool boxcar boxcar used to make other objects 2 bcore, function transfer 0.08
torso pencil1 small, animate and mobile pencil 0 bcore, feature set contrast 0.36
utensil web metal web, used to make food 0 feature set contrast, 0.40
function transfer
vegetable person3 static, inanimate person 0 feature set contrast 0.37
vegetable spoon spoon shaped thing that grows on earth 1 bcore, function transfer 0.50
vehicle body vessel1 vessel made of metal 1 bcore, slot set completion 0.14
vessel1 food concave shaped food in which one can 0 bcore, function transfer, 0.37
put something shape transfer
victim projectionist projectionist damaged by a gun 0 bcore, function transfer 0.25
wheel sitting room circular sitting room 0 bcore, shape transfer 0.25

3.2. Free generation

The free generation of noun-noun combinations consisted of selecting ran-

domly a set of 33 pairs of concepts (the free generation set, which is com-
 Noun-Noun combinations 167

Table 18. The frames used in the experiments

Frame Description
bframe The blend has the same relations of head noun (although the arguments
may differ)
bcore The blend has the same relations and arguments (except those related to
function) of head noun
analogy transfer Transfer all neighbor elements and relations of an element of modifier to
the mapping correspondent of head
function substitution A function from head is substituted by a function of modifier
single differentiating feature Head and modifier differ only on one feature, which is transferred to head
function transfer The head gains a function that was part of the modifier
shape transfer The head gains the shape of the modifier
structure transfer The head gains the structure of the modifier
slot set completion The slots in head that did not have a value are filled with modifier’s
corresponding values
feature set contrast The feature-set in the head are replaced by the feature-set of the modifier

pletely distinct from the tuning set) and using the above described query to
generate new blended concepts. Every frame shown in table 18 was available
to the system so that it could find itself the selection of frames that suited
the highest scores of the fitness function. The optimality constraint weights
were chosen from what we had learned from the previous experiments. In this
case, we wanted to give a central role to the frames (thus giving high value
to Relevance and to Integration), while also allowing a little control to Topol-
ogy, Maximization V.R.50 and Unpacking. The latter received a higher weight
to reinforce inheritance of the input concept’s main characteristics. The val-
ues were: Relevance, 45%; Integration, 30%; Topology, 5%; Maximization of
V.R., 5%; Unpacking, 15%. We also added an integrity constraint of having
at least two frames being accomplished so as to stimulate knowledge transfer.
Apart from these, parameters were equal to those used for tuning.
In figure 39, we show examples of the generation of the “fish tail1 desk”
and “fish spider” blends, with the inputs (“fish tail1”, “desk”, “fish” and “spi-
der”) and the frames that were applied.
In table 19, we show the results achieved. For each pair of concepts, we
show the best result (in terms of the fitness function) of the 30 runs and de-
scribe it textually by enhancing the differences to the head. The use score cor-
responds exactly to the resulting Relevance value. Therefore, a 100% means
that every condition of the query was satisfied and no integrity constraints
were violated. Other values indicate that either some condition was not satis-
fied or that integrity constraints were violated (or both).
168 Experiments

Table 19. Excerpt of the results (average usefulness=78%; standard deviation=35%;

median=100%)

Combination Interpretation use nov Frames

barrel spoon Spoon 1.00 0.09 bcore
bird1 sea A bird shaped sea, with wings, and 1.00 0.67 bcore, shape transfer
head and made of flesh structure transfer,
slot set completion
bird head clothes Curve shaped clothes 0.81 0.12 bcore, bframe,
slot set completion,
shape transfer
cow head torso Conical torso 0.60 0.14 bcore, shape transfer
desk ornament Brown, wooden, ornament one can put1.00 1.00 bcore, function transfer
paper on slot set completion
desk1 spoon bowl Spoon bowl besides which one puts a 1.00 0.31 bcore, function substitution
chair (and is not used to put food in)
engine apple tree Oblong, long and large apple tree 0.43 1.00 bcore, shape transfer
fish spider Spider with fish tail that lives in sea, 1.00 0.69 bcore, function substitution,
but does not make webs structure transfer
fish tail1 desk Thin, triangular desk 1.00 1.00 bcore, shape transfer
flower bloom hammer A spherical hammer 1.00 0.13 bcore, shape transfer
food body part A body part that serves to be eaten 0.35 0.33 bcore, function transfer
herring instrument A silver, fish-shaped (with fin and tail) 1.00 0.87 bcore, shape transfer,
instrument that lives on sea and is not function substitution,
used to play music structure transfer
horse head insect Insect 0.04 0.00 bcore
insect rodent A small rodent 1.00 0.12 bcore, shape transfer
mattress knife A long knife that is on a frame 0.00 1.00 bcore, shape transfer,
function substitution
oak horse A horse that grows on earth, it has a 1.00 0.58 bcore, structure transfer,
trunk and a crown, but keeps its horse function transfer
shape
paper1 chair seat White chair seat 0.50 0.17 bcore,
slot set completion
patient fruit Human shaped, skin-colored fruit that 1.00 0.60 bcore, shape transfer,
is ill function substitution
person5 paper Paper that sleeps in bed 1.00 0.14 bcore, function substitution
person5 stem Stem that sleeps in bed 1.00 0.50 bcore, function substitution
pottery neck A neck made by a human 1.00 0.37 bcore, function transfer
rose bloom desk Desk 1.00 0.42 bcore
sole bird A black bird 1.00 1.00 bcore,
slot set completion
spider legs carriage Carriage 0.67 1.00 bcore, bframe
train building A building with the shape and 1.00 0.54 bcore, function transfer,
structure of a train, and which serves structure transfer,
to transport people shape transfer
utensil pottery Pottery 0.04 0.50 bcore
 Noun-Noun combinations 169

Input Domain 1 Input Domain 2 core(2) Blend

shape_transfer(1)

Input Domain 1 Input Domain 2 bcore Blend

structure_transfer function_substitution

Figure 39. Frames used in the construction of “fish tail1 desk” and “fish spider”

For example, in Figure 39, we can observe that both blends satisfy all
requirements of the query (therefore scoring 100%). If, say, there were no
values for made o f and color, then use would be 75% since two (in eight)
conditions were not satisfied. Another situation could be an integrity con-
straint violation (e.g. “Something cannot be black and made of flesh at the
same time”), which would lead to a penalty (e.g. supposing integrity con-
straint violation penalty was 20%, “fish spider” usefulness value would be
80%). The frames listed correspond to the frames found in the construction
of the best result for each combination.
We notice that every experiment ended satisfying a bcore frame. This is
not surprising considering the query we used, which comprises a set of re-
lations that coincides almost with the bcore frame relations. Still with re-
gard to frames, we can also see that the results used essentially 6 different
frames (bcore, slot set completion, shape transfer, structure transfer, func-
tion transfer and function substitution). A possible explanation may be that
the other 4 were either too specific (single modifying feature) or too generic
170 Experiments

(bframe) to achieve stability in the runs.

Results show that there is no correlation between novelty and usefulness,
which seems intuitively plausible. Yet, use may contrast with our intuition in
some examples (e.g. there is no apparent reason why a “horse head insect”
is less useful than a “rodent insect”) and its explanation is simply that, for
the context we are dealing with, the new object may lack some fundamental
conditions.
Probably because the query is too much centered on the “core” of the
object (every aspect except its function), it may lose its function during the
blend generation, even when it is vital. For example, in blending “herring”
and “instrument”, the result says it is an instrument, but it lost its musical
function, so leading to an empty concept. We also point out to the blends
“train building1” and “bird1 sea”. Both reveal inconsistencies (“a train build-
ing1 is a building that serves to transport people” and “a bird1 sea is a sea
with wings...it is made of flesh”). These inconsistencies may be revealed as
creative if explored from a metaphoric perspective, a very complex compu-
tational challenge although sometimes trivial for humans. Preventing the ex-
istence of these extreme examples depends on adding integrity constraints
(e.g. “something that serves for transportation cannot be made of bricks”) but
these will go against the creative potential of the system.
From the observation of the use scores, it should be clear that the average
of 78% obtained is highly dependent on the specific query and on the spe-
cific knowledge contained in the dataset. If the query was less constrained
(e.g. having just half of the conditions), the score would certainly be higher,
whereas if we added conditions that could not be satisfied within the dataset,
use would never achieve 100%. What these numbers show is that the model
is able to search for the query satisfaction when it is (the knowledge base, the
query and the factory) properly configured, thus providing useful outcomes
for the context in question.
Calculating the measures from Ritchie, we obtain the results in table 20
In comparison with the previous experiments (of horse-bird) shown in ta-
ble 15, we notice an increase of scores for typicality and value (1-4), with
or without considering the inspiring set (11-14). Apart from the inherent dif-
ferences of both experiments, this also reflects a higher control over Divago,
due to the methodology followed (tuning+free generation) and to acquired
experience in weight choice. This experiment also shows a more realistic set-
ting. For example, not all typical items are valued (measure 5) and a few
reinventions were made (9), although these were a very little proportion of
 Noun-Noun combinations 171

Table 20. Ritchie’s (Ritchie 2001) measures results

Measure Value
1 0.543
2 0.563
3 0.782
4 0.781
5 0.778
6 0.344
7 0.786
8 0.786
9 0.036
10 16.000
11 0.513
12 0.831
13 0.500
14 0.781

the results (10). Interestingly, Divago produced almost exactly the same pro-
portion of untypical and valued items as in the previous experiments, with
regard to the whole set of outputs (6). It even increased, if only considering
the untypical items (7).

3.3. Comparison to C 3

We had access to a set of analogous experiments that Costello and Keane did
with C3 . In these experiments, the authors randomly generated 10 pairs of
nouns (e.g. “eagle” and “tulip”) and, for each pair, generated interpretations
for the two possible combinations (e.g. “eagle tulip” and “tulip eagle”). This
gives 20 combinations, for which C3 provided interpretations (e.g. “An eagle
tulip is a tulip that grows on an eagle”). These experiments were intended
to model the creativity of concept combination and therefore it makes sense
to compare them with Divago. However, we cannot do a comparison that
survives subjectivity because both the values referred here (of novelty and
usefulness) and by C3 output (plausibility, informativeness and diagnosticity)
are not aligned in the same perspectives. Surely, C3 interpretations would of-
172 Experiments

ten fail in the use measure suggested by our script, and Divago would not
necessarily do well with C3 constraints, and any of these conclusions would
lead nowhere in terms of saying which one is more creative. We can do a
different, perhaps more interesting, experiment: check if Divago can arrive to
the same results of C3 (thus proving the possibility of achieving the same cre-
ativity, whatever it is); and determine which frames would be needed (would
they have to be different?). First of all, to level both systems in terms of rep-
resentation, we had to allow Divago access to the implicit relations with other
concepts.
In order to check if Divago could find the same results as C3 , we applied
the process described above as tuning phase and found that Divago is able
to achieve the same results with an average error of 2.4 and median 1. This
means that the normal error was either 0 or 1 and so the average was strongly
affected by two outliers, of errors 8 and 10. These latter cases, in which Di-
vago failed, were interpretations that included knowledge from third nouns,
i.e. when there are attributes that do not belong to any of the inputs and come
from other elements in the knowledge base. In the rest, it normally achieved
the same results of C3 . Another remark is that it tended to include knowledge
(e.g. that “an eagle tulip is solid”) that C3 had excluded via the informative-
ness constraint. Whichever one is more correct in this issue, it was also clear
that, by declaring that the diagnostic features of each noun are the features
that differentiate the noun in relation to other nouns (an information that is
actually available in C3 ), Divago could reduce drastically this extra knowl-
edge.
Perhaps the more striking conclusion from this experiment was that Di-
vago could achieve the same results of C3 (with the error just described) with
a very small set of frames. Indeed, only two frames were needed about 85%
of the time: acore (or bcore, depending on whether the focus was the modi-
fier or the head) and analogy transfer. This means that, essentially, C3 picked
one of the nouns (head or modifier), built the combination centered on it -
which means it has the same structure and the same “core” attributes-, and
also transferred the attributes directly related to the other noun. By directly
related we mean attributes with distance 1 in its graph representation. This
seems to indicate that combinations generated in C3 were essentially of the
property type. The other 15% of the results used also the bframe (or aframe,
depending on the focus). The results, representations and C3 results are listed
in appendix E.
 The creature generation experiment 173

To conclude, Divago is able to achieve the same results of C3 by using

a proper set of frames (aframe, bframe, acore, bcore and analogy transfer)
as goals in the search. This means that, if wanting to configure it as a noun-
noun combination interpretation system, only a smaller set of frame com-
binations should be considered, at least for hybrid and property types, and
attention should be paid to other factors, namely to diagnostic features. On
the other hand, considering the other experiments in this book, we conclude
that Divago offers a much larger set of possibilities, without focusing specif-
ically on the linguistics of combinations. In other words, C3 models noun
combinations51 and Divago deals with concept combinations, being more
open to other problem solving situations. We cannot answer the doubt about
the limits of C3 (could it also achieve the same results of Divago, with a
proper configuration?), but it is clear that these are internally very different
systems that tackle the same cognitive problem from different perspectives.

4. The creature generation experiment

We now invite the reader to imagine the following context: a game with a
knowledge base of objects (creatures, physical objects, scenarios, etc.) coded
according to Divago representation. Instead of having a specific object pre-
determined for each game situation, let us suppose only a partial specification
is given (e.g. for situation x, the game needs a f riendly creature that belongs
to the blue team and should have a strength y). Depending on the size of the
knowledge base, on the abstractness of these specifications, and on a com-
petent engine for retrieving these objects, such a game could become more
unpredictable and surprising, which is one of the current challenges in the
area of game development.
We idealized a blending engine for games (Ribeiro et al 2003) that would
fit the context just given, which would partly be a re-implementation of Di-
vago with attention to the specific domain of games and to performance is-
sues, always vital in game development. In order to assess the feasibility of
the idea and have a first insight on the problems involved, we made some
experiments with generating creatures in Divago.
174 Experiments

4.1. Tuning

We built an initial battery of 12 creatures, based on the Magicc

The Gather-
ing game, which comprises hundreds of different creatures, each one with a
strength and defense value pair, a team color, and a mana cost (interpreted by
us as f ood consumption). They could also have functionalities (e.g. protect
another creature) and abilities (e.g. fly). Below, we show an example, the
pa jem angelical (all the creatures used are in Appendix E):

isa(pajem angelical, human) pw(wing, pajem angelical)

isa(pajem angelical, bird) pw(left leg, pajem angelical)
member of(pajem angelical, creature) pw(right leg, pajem angelical)
strength(pajem angelical, 1) actor(strength enhancement,
pw(left arm, pajem angelical) pajem angelical)
defense(pajem angelical, 1) actee(strength enhancement,
pw(right arm, pajem angelical) creature)
food consumption(pajem angelical, 2) pw(head, pajem angelical)
team color(pajem angelical, white) cost(strength enhancement, 0)
color(pajem angelical, human colored) actor(defense enhancement,
pw(torso, pajem angelical) pajem angelical)
made of(pajem angelical, flesh) actee(defense enhancement,
points(strength enhancement, 1) creature)
points(defense enhancement, 1) ability(pajem angelical, fly)
cost(defense enhancement, 0)

For this stage, we had to obtain blends of creatures (to become the targets).
In order to avoid our own bias, we asked another researcher, not aware with
Divago’s inner processes and having little knowledge of Conceptual Blend-
ing, to select randomly pairs of creatures and invent three different combi-
nations for each of them. He chose 14 pairs of creatures, thus making 42
combinations. It was then our task to obtain the set of frames that could help
Divago generate the same set of combinations.
Prior to starting the testing and designing of frames, it was necessary to
check whether the solution actually existed in the search space, as we did
in previous experiments. In other words, given a query with the exact rela-
tions of the target, the mapping applied by the designer52 , no integrity con-
straints (so that, whatever inconsistencies the target may have, it will not be
less valued), and a configuration of 90% weight on Relevance (and 10% on
Integration), Divago should be able, after a sufficient number of generations
(aprox. 30, for the Horse-Bird experiment), to generate the exact same blend.
In such a configuration, the search space has only one maximum, containing
 The creature generation experiment 175

10

9
8

7
Error

4
3

2
1

2
l1

sh ba al3

2
2

2
fu s_d pa te_ e r1

c r r1

m g el _fur urg 3

_g us _ arf1

_m rf3
ve wle un d em l 1

_v awl olta g_ l 3

al
2

2
as _ cr ora _a l ica

ca _p 1
fi e c_c al _ e_c st3

fu u s_ rf_ _a ale r

n
ca raw r_a el e o on

go ma l _a _m n3
go s_d rf_ _ s li ca

r
en c_c _so ia n ant

pa in_ m al _ es_ ha l

sh
ta

st
ic

ite ler

go _ m rf_g _ s on

pa _a sha es_ wle

pa _a li ca l d_ wle

eo

a
al
nt

o
le ou gel

en

ri c ltar prie

ie

i c urio dw
dw
_v je m ge

ar
_s eri c am aw

rio eo s am he

ld ge

ld ge

in rsh bli rge

s
b

gi
e

a
r

ar
m
n

cr

_
n

fi e ur
_
r
_
_a

l _ us
he cr r_a s_

en c_c al _ ia n
s rin
l

l in
e

i
m

io
m

n
_

_c er_ h_

ob
v
r

ie
e m ent aje

h sh

h
je
dw fi e
he aw c i

as
s

f
s
a

f
l_
o
_
p

nt l _p

l
al
h_ le m al _

al
n_
r_

a
t

ic
ic
le
a

a
er
t

w
r
s_ aw

m ar

el
rs
en

le
_v l e
he al

j e ng
l

ng
g
em

n
ur
r

r
e

_a
Blend
s_

_c

_
c

e
el

rio

m
as _e

in

bl
_v
ve
h_

rio
ld

je
bl

bl
fu
el

je
h

ca
as

en
as

as

as

Figure 40. Least possible error theoretically achievable by Divago for each of the
blends (mean values over 30 runs)

either the set of relations of the target or the subset that can be achievable by
Divago.
In figure 40, we can observe that only five of the combinations are com-
pletely contained in the search space. The reason for this apparent fail-
ure is simple when we inspect the combinations that produced bigger er-
ror. Let us analyze one of these combinations, the second combination of
f ield surgeon with pa jem angelical, which has an error of 5. The concept
map of pa jem angelical has been given above and f ield surgeon is repre-
sented as:
isa(field surgeon, clerical) pw(head, field surgeon)
member of(field surgeon, creature) pw(left arm, field surgeon)
strength(field surgeon, 1) pw(right arm, field surgeon)
defense(field surgeon, 1) pw(torso, field surgeon)
food consumption(field surgeon, 2) pw(left leg, field surgeon)
team color(field surgeon, white) pw(right leg, field surgeon)
color(field surgeon, flesh colored) actor(healing, soltarian priest)
made of(field surgeon, flesh) actee(healing, creature)
size(field surgeon, medium size) points(healing, 1)

The designed combination for f ield surgeon|pa jem angelical2 was:

176 Experiments

pw(pajem angelical right arm, field surgeon)

member of(field surgeon, creature)
pw(pajem angelical torso, field surgeon)
strength(field surgeon, 2)
pw(pajem angelical wing, field surgeon)
defense(field surgeon, 1)
pw(field surgeon left leg, field surgeon)
food consumption(field surgeon, 1)
actor(defense enhancement, field surgeon)
team color(field surgeon, white)
actee(defense enhancement, creature)
points(defense enhancement, 2)
made of(field surgeon, flesh)
color(field surgeon, human colored)
pw(field surgeon head, field surgeon)
cost(defense enhancement, 1)
pw(pajem angelical left arm, field surgeon)
moves(field surgeon, jumping)
As we can see, there was a scrambling of all the numbers involved
(strength, defense, food consumption, points and cost). This would be no
problem if the substitution was consistent with the projection mechanism of
Divago, but this was not the case: sometimes the number 1 becomes pro-
jected to 2 (both projection of strength), sometimes to 1 (in de f ense), the
case becomes even more complicated because 0 projects also to 1 (in cost).
The problem we are raising is that, by definition, a selective projection can
have one and only one projection for each concept in the input concept maps,
thus even if, by the mapping function, 1 is mapped to 0, 1 and 2 (which
wouldn’t be possible anyway with our 1-to-1 structure alignment algorithm),
it can only be projected as one of these in the blend. This now seems to us
a serious limitation. Another problem in this blend is that there is one com-
pletely new concept, jumping, which did not exist in any of the inputs. The
reasoning followed by the designer was that, since the new creature has only
one leg, then it can only move by jumping. This was important information,
since it gave rise to a rule in a knowledge base. After determining the least
possible error, we may proceed to determining the frames.
We can see that some of the previous frames (aframe, bframe, acore,
bcore) were simplified with parameters (frame(X) and core(X), respectively),
although maintaining the same reasoning. In figure 41, we show the best per-
formance achieved and in figure 42 we show the difference compared to the
best possible result. The weight configurations remained the same (90% Rel-
 as
h_
as ele
h_ m
as ele en
as h_e me tal_
h_ le n pa
as ele me tal_ jem
h_ m n pa _
as ele en tal_ jem ang
Missing Error

core(X)
h_ m tal pa _ el

0
1
2
3
4
5
6
7
8
9
10
el en _v jem an ica

shape(X)
frame(X)
e ta o g l
as me l_v rac _an elic 1
he nta o iou g al
creature(X)

fightAttr(X)
as s_ l_v rac s_ elic 2
he cra or iou ba al3
as s_ wl ac s_ boo
as he cra er_ ious bab n1
he s_ wl as _ oo
as s_ cra er_ h_ bab n2
he cr wl as ele oo

attr transfer(L, X)
as s_ awl er_ h_ me n3
he cra er as ele nt
s_ w _th h_ m al1

function transfer(X)
shape transfer(S, X)
ca cra ler_ und ele ent
shape amputation(L)
ve w th e m al2
ca _c ler_ und ring ent
ve ra th e _ al3
ca _c wle und ring gian
ve ra r_ e _ t1
en _c wle sol ring gia
_ ra r_ tar _ nt2
en vec wle sol ian gia
_v _c r_s tar _p nt3
en ec ler ol ian rie
_ _ ic ta _ s
en ve cle al_ ria prie t1
_ c_ ric ca n_ s
en vec cle al_ ve_ prie t2
_ _c ric ca c s
en vec ler al_ ve_ raw t3
_ _c ica ca c le
fie vec ler l_s ve_ raw r1
ld _c ica am c le
from input X

fie _su ler l_s ite raw r2

ld rg ica am _h ler
fie _su eo l_s ite ea 3
ld rg n_ am _h ler
_s e pa it e 1
o e a
fu urg n_p jem _h ler
rio eo a _ ea 2
a
fu us_ n_p jem ng ler
rio d a _ el 3
a
fu us_ wa jem ng ical
r r
fu iou dw f_fi _an elic 1
rio s_ ar el g al
d
fu us_ dw f_fi _s elic 2
rio d ar el ur al
d
fu us_ wa f_fi _s geo 3
rio d rf_ el ur n
w
go us a go d_ ge 1
b _d rf bl su on
go lin_ wa _go in_ rge 2
b m rf bl m on
go lin_ ars _go in_ arsh 3
bl ma ha bli ma al
i l
Table 21. List of frames used in the creature blends

n 1
go n_m rsh _as _m rsh
b a al he a al2
go lin_ rsh _as s_ rsh
bl ma al_ he cra al3
i

Figure 41. Best possible + frame results (mean values)

n r a s w
Best results+goal frames

go _ s s _ l
b m h h c e
pa lin ars al_f es_ raw r1
je _m h ie c le
pa m_ ars al_f ld_ raw r2
je an ha iel sur ler
pa m_ ge l_f d_ ge 3
l s
The creature inherits from X the shape S

pa jem ang ica ield urg on1

je _ el l_f _s e
pa m_ ang ica urio urg on2
je an eli l_fu us eo
pa m_ ge ca rio _d n3
The creature inherits a function from input X

je an lica l_fu us wa
m g l_ r _ rf
_a eli g io dw 1
ng ca ob us_ ar
el l_g lin dw f2
ic o _m a
al bl
_g in ar rf3
ob _m sh
The creature inherits the overall shape from input X

lin ar al1
_m sh
ar al2
s
The creature contains the core (all attributes) of input X

Blend
The creature maintains the same set of relations of input X
The creature generation experiment

evance, 10% Integration). In Appendix E, we list the generated blends.

Tuning
Least Error
The creature inherits from input X the set of attributes in list L
The creature inherits the fight attributes (strenght and de f ense)

AverageTuning
AverageLeastError
The name of the creature is the same as the name of the creature X

given, the search space became much too convoluted and, in order to accom-

different local maxima throughout the 30 runs. An example of such incom-

find a set of stable combinations of frames. Whatever combination of frames
177

Except for a few situations, it was not difficult to find a combination of

Since we tried to avoid tailoring the frames to the specific blends, for some

plish some frames, Divago had to drop others, eventually achieving many
frames that would give us the best result or a result that was very close to it.

(namely the ones with error bigger than 2, in figure 42), it was difficult to
The creature has not any of the shape items(arms, legs, etc.) in list L
178 Experiments

10
9
8
7
6
Error

5
4
3
2
1
0

2
l1

l3

fi e urg l 2

2
2

r3

c r r1

pa _an al _ _su r3

us arf1

_m rf3
l1

l3

al
n_ nt2

2
_a ic a

er
ite er2

_a li ca iou eon
on

_ m rf_g _ su n1

_m on3
a

al
le

sh
ta

_c er_ erin nta

ia ies t

st
ic

le

le
ic

a
al
em hea

eo
_v wle ltar _gi a

sh
h_ abo
el

ou gel

aw

aw
en

av prie

dw

w
l

el

ar
w

e
fu rge _sa _ he

rg
e
ng

pr

_d
g

rg

ar
ra
m

ca raw hun l em
n

cr

s_
n
b

_
n_
g
_a

c
s_

ite

u s ar f _ _ a

s_

_a sha es_
s
e_

l in
el

ia

ld
ta e m

u s war ld_

rio
ri c am

li n

he
h_

ob
Blend
_v l eri ltar

fie

r
j e ma ash
je

fu

fu
e

ob
i

as
aj

as

as
c

_d pa j
as ent _pa

_g
c

l_
i

l_
o

o
a

l_

f
h_ e m l _p

l_

f_

_ m hal _

_
r_

r_

s
r

al

al
ec ri ca
vo

_
ca

al
r_
_c r_t

ic
l

le
ta

on

ic
sh
l
he al _

el
w

r
w

s_ aw

ge

el
s
en

en

le

le
le

w
d

ar

ar

ng
ra

ng
w

_c

_c
c

_
as e m

_d
a
r
c

us
a

_
r

u
ec
s_

s_

ec
cr

in
_s
em
el

el

rio

rio

m
in

in

bl

m
he

_v
ve

ve
h_

h_

rio
ld

je
bl

bl

go
fu

je
el

he

ca

en

en

fi e
as

en

go

go

pa

pa
fu
as

as
as

Figure 42. Efficency of Divago: difference to best possible values (mean values)

patible frames are the frames for shape amputation(List) and shape(X ) (in
table 21). The former removes a piece of the creature, while the latter tries to
make it as a whole.

4.2. Free generation

In order to see the capacity of Divago of generating novel creatures, we ap-

plied several different configurations and creature combinations. Given the
game context, this time we want to apply elaboration and produce a visual
output. Due both to copyright obligations and to the availability of a set of
three completely defined 3D creatures (a werewolf, a dragon and a horse, see
figure 43), the content of the knowledge base was changed to include only
these latter objects, coded in a similar manner as the creatures from Magic c

The Gathering used above.

By changing the knowledge base and keeping the frames and rules ob-
tained, we may verify that Divago has some degree of versatility. In fact, since
we maintained the same kinds of relations in the concept maps, only a few
rules and integrity constraints had to be created. We added rules for stating
 The creature generation experiment 179

Figure 43. The creatures available in the knowledge base

that something that has wings should have the ability to fly, something with
an odd number of legs should get a wooden leg in the missing connection and
it should move by jumping. Another rule calculates the mean when there are
two different numerical values for the same attribute (and replaces them with
this mean). We also added a rule stating that, when something is dangerous,
very strong and very large (strength > 5 and size > 3), then it should get an
ogre head and lose the original one. The new integrity constraints state that
a creature should not have two different values for the same attribute, they
should be symmetric, they should not have two heads, two torsos, or two
members in the same place (e.g. two le f t arms)53 .
We only applied a single query throughout this experiment. To determine
this query, we analyzed the history of frame combinations used to build the
42 creatures in the previous stage. We concluded that each one had two or
three abstract frames such as creature(X), frame(X) or core(X). Since core(X)
can be too specific, we decided to have creature(X) and frame(X) (the former
forces to only have one creature name in the blend, the latter to follow its
relational structure). There is also some regularity in the transfer of shape
parts from each of the inputs, so we decided to have shape transfer(E1, Y),
shape transfer(E2, Y), {E1 \= E2}, with X different from Y. Thus, the query
is:
creature(X ), f rame(X ), shape trans f er(E1,Y ), shape trans f er(E2,Y ),
{E1\ = E2, X \ = Y }.
A good blend would therefore consist of the structure of one of the crea-
tures with at least two of the shape parts of the other creature. In table 22, we
show, for each pair of creatures and a weight configuration, the median results
of novelty and usefulness obtained, as well as the scores for the best blend.
The weight configurations consist essentially of the ones used in previous
experiments.
180 Experiments

Table 22. Results of creature combinations

Combination Weights nov use Best blend
I T MVR U R (nov/ use)
horse werewolf 34 16 10 4 36 0.50 1.00 0.56/1.00
horse dragon 34 16 10 4 36 0.31 1.00 0.25/1.00
werewolf dragon 34 16 10 4 36 0.50 1.00 0.62/1.00
horse werewolf 10 0 0 0 90 0.86 1.00 0.86/1.00
horse dragon 10 0 0 0 90 0.69 1.00 0.57/1.00
werewolf dragon 10 0 0 0 90 0.74 1.00 0.65/1.00
horse werewolf 20 0 0 0 80 0.86 1.00 0.86/1.00
horse dragon 20 0 0 0 80 0.87 1.00 0.75/1.00
werewolf dragon 20 0 0 0 80 0.76 1.00 0.65/1.00
horse werewolf 30 5 5 15 45 0.83 1.00 0.71/1.00
horse dragon 30 5 5 15 45 0.37 1.00 0.50/1.00
werewolf dragon 30 5 5 15 45 0.59 1.00 0.65/1.00

In these experiments, almost every result entirely satisfied the query, thus
giving a value of 100% for each one (the mean was 0.98 with a standard devi-
ation of 0.05 corresponding to a few outliers). With respect to novelty, we can
observe the variability of the results with the weight configurations. Indeed,
when there is a focusing on Relevance and Integration, the system runs away
from typicality, which is understandable when we analyze the used frames.
They favor the use of knowledge from both inputs, without significantly fa-
voring one input over the other. When doing so, as also verifiable in the other
experiments, novelty tends to increase. When, on the contrary, the frames fa-
vor one of the inputs (e.g. the Pegasus is a horse), then the typicality will
tend to increase. When adding other optimality principles such as namely
Topology and Unpacking, we can notice a decrease of novelty, although also
achieving the 100% solution. This results from a heavy weight on Relevance,
but also a preference for those blends that, although accomplishing the goal
frames, respect the other principles as much as possible (Topology and Un-
packing particularly favoring similarity to inputs). From applying Ritchie’s
measures, we obtain the table 23, where we can see that the average and ratio
of typicality has lowered slightly in comparison to the previous experiments
(measures 1 and 2). The average and ratio of value is, of course, 100% given
that every result entirely satisfied the query (3 and 4). This fact implies also
maximum values in other measures (5, 7, 12 and 14). Again, this demon-
 The creature generation experiment 181

Table 23. Ritchie’s (Ritchie 2001) measures results

Measure Value
1 0.343
2 0.333
3 1.000
4 1.000
5 1.000
6 0.667
7 1.000
8 2.000
9 0.000
10 N/A
11 0.308
12 1.000
13 0.333
14 1.000

strates the difficulties both in determining the value of something as well as

in comparing it with other experiments. The ratio of untypical and valuable
results, a very important measure for creativity, was raised to 0.667 (remem-
ber that in previous experiments, it was rounded to 0.333), expectable given
that everything is now maximally valued. The rest of the measures confirm
the same conclusion: Divago was able to satisfy the value criteria for every
combination, it did not reinvent any of the inputs and it was able to produce
some proportion of results with low typicality (i.e. high novelty).
The generation of these 3D images was made by an interpreter devel-
oped in collaboration with other researchers (Ribeiro et al 2003). It receives
the concept maps generated by Divago and produces a “wavefront obj” file,
which describes the 3D image. The several parts of the creature were coded
separately (e.g. horse back leg, ogre head) and placed together according to
the concept map. To give an idea of the creatures generated, we now show
some examples (the rest in appendix E). In figures 44, 45 and 46 we show the
images of the best blends found in configurations 1, 2, 3 and 4, respectively.
In order to give an insight on the range of generated creatures, we also
show the worst results. In these, we can see that they either lack one member
(e.g. a wing), they have more than one member in the same point (e.g. a
182 Experiments

Figure 44. The best blends for horse|dragon (nov=0.25), horse|werewolf (0.56) and
werewolf|dragon (0.62)

Figure 45. The best blends for horse|dragon (0.37), horse|werewolf (0.86) and
werewolf|dragon (0.65)

horse|werewolf with four back legs) or a pirate leg. See figure 48.
We left the visualization of these creatures to the end of this section to
prevent the reader from placing excessive importance on the images. Indeed,
there is much more behind each of these creatures, namely characteristics
such as the abilities, their strength and de f ense values and so on. Therefore

Figure 46. The best blends for horse|dragon (0.75), horse|werewolf (0.86) and
werewolf|dragon (0.65)
 The creature generation experiment 183

Figure 47. The best blends for horse|dragon (0.50), horse|werewolf (0.71) and
werewolf|dragon (0.59)

Figure 48. Some worst blends for horse|dragon (nov=0.73), horse|werewolf (0.44
and 0.60) and werewolf|dragon (0.67)

their novelty, as reported in the captions, may have been affected by these
non-visual characteristics.
The positive conclusion from these experiments was that Divago, as a gen-
erative model, can enhance the dynamics of a game environment. Suppose
that it is allowed to blend not only creatures, but also scenarios, physical ob-
jects, behaviors, and so on. Even more, with appropriate frames, it is theoreti-
cally possible to blend creatures and scenarios (with blends that, say, transfer
the color or texture of a scenario to a creature, or the function of an object),
which would considerably potentiate the possibilities of the game.
Although these experiments were stimulating as a motivation for develop-
ing a game, they also revealed some problems that need to be solved:
– The majority of the mappings used in this experiment were not made by the
Mapper, they were hand coded, first because the Mapper is easily fooled by
the representation simplicity (e.g. it can map back leg to right arm because
both are connected to the creature with the same relation, pw), and second
because some of the mappings are not based on structure alignment. The
above situation of the scrambling of numbers is a good example.
184 Experiments

– The projections are too restrictive, which prevents them from achieving a
lesser error to the targets. Perhaps the game engine should consider multi-
ple projections for the same concept map element.
– Divago is extremely slow in generating each creature. This is acceptable
for a Prolog based prototype, but not for an on-line system. For this reason,
the game engine must be developed in a computationally faster and lighter
setting.

5. The established Blending examples

One of the contributions of this work is a computational model of Conceptual

Blending and therefore it is fundamental to validate it with a set of examples
recognized in literature as being conceptual blends. In section 3.2 and in Ap-
pendix B, we present what we call the established Blending examples. These
are Blending case studies that appear in literature and that we think should be
considered when building a computational model of Conceptual Blending.
Although more examples could be included, we restricted ourselves to the
ones sufficiently specified (they should at least discriminate all the mappings
and elements from the inputs) and which considered only two input spaces
(there are many examples with multiple input spaces). Moreover, they should
be considered “conceptual blends” somewhere in their description, to prevent
any subjective evaluation from our side.
In table 27, we enumerate the examples, as well as their characterization.
We believe that they are representative of a number of situations that have
been approached over the last few years in the main CB reference literature.
In order to provide Elaboration, a few rules were added to the generic domain,
namely the movement laws rule (as presented in section 7) and rules stating
common sense implication (e.g. “When an x and a y are married, they form a
couple”). The Mapper was not used because the mappings for each example
were already given in the literature.
From each of these examples, we extracted the input domains, the blend,
the generic space and the mapping. Normally, all these were directly avail-
able in tables or diagrams. The special cases were “Computer Desktop” and
“Computer Virus”, which were completed with some common-sense knowl-
edge inserted by us.
For each of the examples in table 27, the goal of these experiments was
 The established Blending examples 185

Table 24. 12 examples of Conceptual Blending found in literature

Name Typology
The Riddle of the Buddhist monk Mirror network, Topology preserving
CEO boxing fight Single-scope network
Gun wound Nominal compound, Double-scope network
Kant debate Double-scope, Mirror network,Pattern Completion
Land yacht Nominal compound, Analogical
Trashcan basketball Double-scope
Computer desktop Double-scope,Metaphorical
Computer virus Double-scope,Category metamorphosis
Same-sex marriage Double-scope, Category metamorphosis
Sandwich counterfactual Counterfactual blend, Single-scope
“Mabel is the daughter of Paul” XYZ blend,Single-scope
Pet fish Nominal compound

to understand to what extent Divago was able to achieve the “correct” blends
(the targets). Unlike the previous experiments, we will not focus on novelty
and usefulness as the goal here is not to assess the creativity of the system,
but to find how competent it is in being a model of CB, i.e. in minimizing the
error to the target. These targets correspond to the blends described in each
example in literature. In table 25, we show the target for the “gun wound”
blend as well as a (hypothetical) blend with error 0.75. It has 2 relations that
do not belong to the target (“actee(shoot, target)” and “result(shoot, result)”)
and misses one target relation (“actee(shoot, human)”), yielding a sum of 3.
Since the size of the target is 4, we have an error of 3/4 (=0.75). Please no-
tice that this error measure (which follows the same reasonings described for
calculating distance in the Horse-Bird experiments, in section 2) potentiates
error values that may be higher than what intuition would say. For example,
if a projected relation r erroneously replaces another one, this will count as
two (one for delete, another for insert) instead of one (i.e. there is “only one”
wrong relation). This doesn’t mean we should divide by two the estimated
errors, but we must take this into account when quantitatively analysing the
errors found.
As in the Horse-Bird experiments, we divide the experiments into two
different stages: isolated constraints and combined constraints. In the former,
we will be able to watch the behavior of each isolated optimality constraint of
Divago w.r.t. each of the examples. In so doing, it is possible to observe which
of these constraints is able to achieve minimum error in the blend generation.
The results will also be useful for the last stage of the experiments, in which
186 Experiments

Table 25. Target blend for the “gun wound” example (left) and a blend (right)
actor(shoot, agent)
actor(shoot, agent).
actee(shoot, target)
actee(shoot, human).
error=0.75 −→ means(shoot, gun)
means(shoot, gun).
result(shoot, wound)
result(shoot, wound).
result(shoot, result)

Table 26. An informal description of the frames used

Frame Description
aframe The blend has the same relations of input space 1 (although the
elements may differ)
aframe The blend has the same relations of input space 2 (although the
elements may differ)
aprojection The blend has the elements of input space 1
bprojection The blend has the elements of input space 2
analogy transfer [Part of] the blend results from the transfer of all neighbour
elements and relations of mapped elements of input space 2
to their counterparts in input space 1
role transfer [For a noun-noun compound blend] the head is projected to
a role element of the modifier (e.g. for “gun wound”, “wound”
is projected to “result” in the blend)
Debate The blend has all the relations expected for a debate scenario
(see example 3)
day compression All temporal elements of both inputs become referent to the
same day
head transfer [For a noun-noun compound blend] all relations and elements
connected to the head (input space 2) are projected untouched
to the blend

we will combine constraints according to the results obtained previously. We

will not focus on the principles of Web, Maximization and Intensification of
Vital Relations. The reasons are now obvious: Web is not considered indepen-
dent in our implementation, thus it wouldn’t make sense to test it in isolation;
we are not considering the compression role of Vital Relations, which is after
all their reason of existence, thus the results of Maximization/Intensification
of Vital Relations wouldn’t produce conclusions that we could confront with
their own corresponding theory.
We also followed the methodology for the noun noun experiments, with a
first, tuning, step to obtain the frames. The set of frames achieved is listed in
table 26 and we show their code in appendices B and E. The table 27 shows
the goals used in each of the examples.
 The established Blending examples 187

Table 27. The goals used for the Relevance principle

Goal
Example Frames Relations
Buddhist monk day compression meets(monk1, monk2)
CEO boxing aframe, bprojection
Gun wound aframe, role transfer
Kant debate bprojection, debate
Land yacht aprojection,
head transfer
Trashcan basketball aframe,
analogy transfer
Computer desktop aframe, bprojection,
analogy transfer
Computer virus aframe, head transfer
Same-sex marriage aframe same sex(person1, person2),
married with(person1, person2)
Sandwich counterfactual
Mabel is the daughter of bframe, aprojection
Paul
Pet fish aframe, bprojection

In table 27, we can observe some regularity in the choice of frames. Nor-
mally, there is at least one generic organizing frame (e.g. aframe, bprojec-
tion), which establishes the general structure of the blend (aframe makes the
blend maintain the relations of input 1; bprojection makes the blend main-
tain the elements of input 2). Then, there may be other frames that can be
transforming (e.g. analogy transfer, head transfer, day compression) or or-
ganizing (e.g. role transfer). Finally, pattern identifying frames like debate
are used in specific situations. A final remark concerns the “sandwich coun-
terfactual” example. As referred to in the creatures experiment, Divago is
technically unable to reach the target when a 1-to-many mapping is necessary,
as happens with this blend, so we removed it from the rest of the experiments.
Apart from Relevance, no other Optimality principles demand special con-
figuration concerns, therefore their application depends exclusively on their
weight being higher than zero.

5.1. Experiments with isolated principles

The intention of this part of the experiment is to find the dominant principles
of each blend within the first order i.e., it gives us the principle that seems to
188 Experiments

Optimality Principles in Isolation

3.5

3
Buddhist Monk
Computer Virus
2.5
Gun Wound
Error

Land Yacht
2 CEO Fight
Clinton Lewinsky
1.5 Same-sex Marriage
Kant debate
1 Trashcan Basketball
Computer Desktop
0.5 Pet fish
Mabel is the...

0
Relevance Integration Pattern Topology Unpacking
Completion

Optimality Principle

Figure 49. The median of the error, for the optimality principles in isolation

be immediately prevalent in the blend (within the second order, we would see
pairs of principles that seem to be prevalent, and so on), thus giving a first
classification for our blends. This latter idea can become even more precise
when considering the behavior of all (isolated) principles for each blend and
comparing them as a set, so, instead of comparing each value individually,
we can compare a sequence of values.
In figure 49, we present the overall results of each of the optimality princi-
ples. An immediate conclusion can be drawn from this graph: the Relevance
principle has consistently smaller error than any other principle; the “CEO
Fight” blend beats the record of maximum error in two of the principles. This
confirms the results of previous experiments and highlights the importance
of this principle. On the other hand, it demonstrates that almost none of the
blends have a naturally inherent tendency for the other principles in isola-
tion (at least, as we implemented them). Furthermore, since the Relevance
principle is configured differently for each blend according to an intuitive,
 The established Blending examples 189

Error Optimality Principles in Isolation

3.5

Buddhist Monk
3
Computer Virus
Gun Wound
2.5 Land Yacht
CEO Fight
2 Clinton Lewinsky
Same-sex Marriage
1.5 Kant debate
Trashcan Basketball
1 Computer Desktop
Pet fish
Mabel is the...
0.5

0
Relevance Integration Pattern Topology Unpacking
Completion
Optimality Principles

Figure 50. Reduced line graph corresponding to the error of isolated optimality prin-
ciples

trial-and-error, choice of goals, it seems to indicate that there is no generic,

context-independent, principle that can lead to an exact solution. It is also
important to say that, since Relevance demands a specific configuration, its
value may vary immensely according to the choices of the goals used and so
the reader should retain this aspect while interpreting these graphs.
A more practical conclusion from figure 49 regards the graphical repre-
sentation itself. Although using bars seems to be correct given that we have
no scale in the x axis, it does not simplify the task of understanding and or-
ganizing the several kinds of blends we may be considering. For this reason,
we decided to represent it as a line graph (fig. 50).
Analyzing the results paying attention to the typology presented in table
27 was our first concern for this stage. However, we found no particularly
revealing patterns in the results. There may be many different explanations
for this, but the most salient one suggests that there may be no correlation be-
tween any of the principles and the given typologies. For example, a double-
scope blend may be Topology preserving and another one may demand strong
Integration, a mirror network may also preserve (or not) Topology. For this
190 Experiments

latter case, even if it does preserve Topology, the value of this principle may
not have to be very high (i.e. in a mirror network, there is a priori topologi-
cal correspondence between the input spaces, so the preserving effort may be
low). Many other arguments that testify to the complexity of blends could be
given, eventually ending up in the uncertainty of the typology itself (e.g. the
difference between single and double-scope can become extremely subtle).
On the other hand, we found clear patterns in the choices of goal frames
for the Relevance principle. Both single-scope blends achieved exactly the
target with the pair xframe and yprojection (being x and y either “a” or “b”).
This completely agrees with the idea of single-scope - the elements of one of
the inputs are organized according to the frame of the other. Nominal com-
pounds also show a pattern: there is a “projection” of one of the inputs, the
one that coincides with the focus of the compound. The exception is “gun
wound”, which is also double-scope. Double-scope examples normally de-
mand more specific frames (e.g. debate, analogy transfer, etc.) or specific re-
lations (e.g. same sex(person1, person2)) and were less consistent in reach-
ing the exact target. This confirms that “in a two-sided network [i.e. double-
scope] (...) it is necessary to use a frame that has been developed specifi-
cally for the blend and that has central emergent structure. (...) In two-sided
networks, then, we expect to see increasing competition between optimality
principles and increasingly many opportunities for failure to satisfy them”
(Fauconnier and Turner 1998).
Considering a qualitative evaluation based on similarity in terms of the
shape of the graph and of the principles that yield smaller error, we found
four different groups of blends: Group 1 (“Same-sex marriage”, “Computer
Desktop”, “Pet fish” and “Gun wound”); Group 2 (“Computer Virus”, “Kant
Debate”, “CEO fight” and “Trashcan Basketball”), Group 3 (“Mabel is the
daughter of...”) and Group 4 (“Buddhist Monk” and “Land yacht”). These
are shown in figures 51, 52, 53 and 54.
Except for “Computer Desktop”, the blends in Group 1 have Relevance
yielding the smaller error being followed by the Integration and in similar
proportion by Pattern Completion. “Computer Desktop” was the only exam-
ple that we weren’t able to find a good set of goal frames for. As a result,
isolating Relevance yielded an extremely large error. Thus, the reason for in-
cluding it in this group lies on the other principles. We can see that, as in the
other examples of this group, Integration yields the smaller error, followed by
Pattern Completion. Topology and Unpacking show a less constant pattern,
although all falling within a small error range.
 The established Blending examples 191

Group 1

2.5

1.5
Same-sex Marriage
Error

Computer Desktop
Gun Wound
Pet fish
1

0.5

0
Relevance Integration Pattern Completion Topology Unpacking
Optimality Principles

Figure 51. Group 1

Group 2

2,5

Kant debat e
Error

Comput er Vir us
1,5
CEO Fight
Tr ashcan Basket ball

0,5

0
Relevance Int egr at ion Pat t er n Complet ion Topology Unpacking

Optim ality Constraints

Figure 52. Group 2

192 Experiments

Group 3

2,5

1,5
Error

Mabel is t he...

0,5

0
Relevance Int egr at ion Pat t er n Complet ion Topology Unpacking

Optim ality Principles

Figure 53. Group 3

Figure 52 shows Group 2, which consists of blends that have in Pat-

tern Completion the second smaller error and in Integration and Topology
the two highest errors. Another interesting remark is the high similarity be-
tween the results of “Computer virus” and “Kant debate” (except for Rele-
vance which, as said previously, may vary according to goal configuration).
Group 3 has only a single example. As we see in Figure 53, there is a pattern
of zero error in Relevance followed by stabilization around a specific error
value. Every principle (except Relevance) got an error of 2. After analyzing
the results carefully, we understand that: a) it is based on a single relation
(daughter o f (mabel, paul)); b) being so, the target is only achieved when
bframe and aprojection are achieved simultaneously, which does not happen
consistently except in Relevance. In Integration, we observed that the system
gives a higher score to the accomplishment of bframe alone, instead of its
combination with aprojection. This agrees with our intuition for Integration
presented in section 6, in which we argue that, when the blend is totally cov-
ered by a single frame, its Integration value should be stronger than when it
is totally covered by two different frames.
The examples that have Topology and Relevance as the main principles
were gathered in Group 4 (fig. 54). In fact, particularly for the “Buddhist
 The established Blending examples 193

Group 4

2,5

1,5
Error

Buddhist Monk
Land Yacht

0,5

0
Relevance Int egr at ion Pat t er n Complet ion Topology Unpacking

Optim ality Principles

Figure 54. Group 4

Monk”, this agrees with the analysis of Fauconnier and Turner (Fauconnier
and Turner 2002: 45), who stress the role of Topology as being fundamental
for this example. The results of “Land yacht” were unexpected and some
further analyses revealed that the target is topologically very similar to one
of the inputs (the “land”) and so the highest value in Topology may get close
to the target, particularly when few elements are projected from the other
input (the “yacht”), thus having, in the blend, a copy of the “land” domain.
Moreover, since this example had difficulty in achieving small errors (except
for Relevance), the salience of Topology is probably magnified.
There seems to be no specific pattern underlying the groupings found. We
analyzed issues like concept map size, emergent structure and difference to
input domains, but still no patterns were found. In the next section, in which
we will apply combinations of weights to each of the groups, it will be possi-
ble to check if the same grouping tendencies are maintained. If this happens,
then we will have more evidence of the meaningfulness of the groupings
made.
194 Experiments

Table 28. Weights used in the experiments with combination of principles

Weights
Combination (I, PC, T, U, R)
Relevance+Integration (Group 1) (50, 0, 0, 0, 50)
Relevance+Pattern Completion (Group 2) (0, 50, 0, 0, 50)
Relevance+Integration+Topology (Group3) (33.3, 0, 33.3, 0, 33.3)
Relevance+Topology (Group 4) (0, 0, 50, 0, 50)
Combination 2 (20, 20, 20, 20, 20)
Combination 3 (38, 0, 18, 4, 40)
Combination 4 (21, 0, 21, 5, 53)

5.2. Experiments with combination of principles

The second part of these experiments is dedicated to combining the optimal-

ity principles in different ways. We made four different combinations: the
two principles with least error from the previous stage of these experiments;
all principles with equal weight (Combination 2); application of two sets of
weights derived from Horse-Bird, normalized to exclude MVR (Combina-
tions 3 and 4, resp.). In Table 28, we show the weight configurations used.
We organized the results by maintaining the groupings already discussed
and, for each blending example, we also added the best value of error obtained
in the previous section in order to better understand the evolution.
The first observation is that combination 2 tends to obtain the worst re-
sults, meaning that adding all principles with equal weights reveals unpro-
ductive, supporting the idea that each blend may result from different combi-
nations of these competing pressures. Another immediate observation regards
the difficulty in approaching the target, i.e. in comparison to the best so far
achieved (with the isolated principles) the error only becomes smaller than
this best in the examples of “Kant debate” and “Trashcan basketball” (both
when combining Relevance with Pattern Completion), while in “Gun wound”
and “Buddhist monk”, the first combination reaches the best solution error.
For all the others, no combination brought better results. This raises perhaps
the most important question for this experiment: is Relevance (or the way we
configure it) sufficient to achieve the best blend? From all the experiments
done so far, the answer turns out to be “yes”. Indeed, with a proper goal
choice, it is in principle possible to find the target, independently of its com-
plexity. And this takes us to the perspective of classifying blends by frames
 The established Blending examples 195

Group 1

2,5

1,5
Same- sex Mar riage
Error

Gun Wound
Pet f ish
Comput er Deskt op
1

0,5

0
Best Relevance+Int egrat ion Combinat ion 2 Combinat ion 3 Combinat ion 4

Figure 55. Combinations of weights applied to Group 1

rather than by optimality principles.

It is also observable that the four groupings found in the previous section
are generally consistent with the new results, although the groupings may
seem dubious at points (e.g. we could equally imagine the interchange of
some examples from Groups 1 and 2 or from Groups 3 and 4). Therefore,
we may now find two distinct major groups (let us call them clusters), the
first one comprising groups 1 and 2. In this cluster, Combination 2 normally
gets the highest error, followed by Combination 4. Except in Group 2, there
is also a tendency for obtaining the same value for Combination 3 and the
“two best principles” combination. The other cluster comprises groups 3 and
4, in which Combination 4 seems to acquire much better results than Com-
binations 2 and 3. This may mean that the blends in question demand less
Integration (which can be confirmed in the previous section), since Combi-
nation 4 differs from Combination 3 essentially in the weight of Integration.
Still in this cluster, we can also see that one example, the “Buddhist monk”, is
very stable in the obtained errors, getting worse results only when the weight
of Relevance is shared with all others (Combination 2).
A final and more practical point regards the computational performance
of the system. Given the complexity of some blends and our experiment re-
196 Experiments

Group 2

3,5

2,5

2 Comput er Virus
Error

Kant debat e
CEO Fight
1,5 Tr ashcan Basket ball

0,5

0
Best Relevance+Pat t er n Combinat ion 2 Combinat ion 3 Combinat ion 4
Complet ion

Figure 56. Combinations of weights applied to Group 2

quirements (30 runs for each blend, for each configuration), our search engine
needed a considerable amount of time in some cases. We used an Intel Pen-
tium IV  at a speed of 2.4 GHz, which needed sometimes four to five hours
to find a solution (only in the most complex cases like “Computer Desktop”
with Combinations 2, 3 or 4). At best, it took 2-5 seconds to find a result.
These values can get much lower after an optimization of the system, but it
will find it hard to become fast enough to enable comparison with the per-
formance of our own cognitive system, particularly if taking into account the
extremely large amount of background knowledge that we are able to cope
with.

5.3. Some considerations

Given the complexity involving several aspects of Conceptual Blending and

the extremely wide range of situations considered, the Blending mechanism
of Divago is far from an exhaustive model. Its initial and most fundamen-
tal motivation was to be the Bisociation module for Divago. Thus, its scope
can be considered very limited in comparison to what we can find in all the
CB literature. This fact does not inhibit us from understanding how much
 The established Blending examples 197

Group 3

2,5

1,5
Error

Mabel

0,5

0
Best Relevance+Topology Combinat ion 2 Combinat ion 3 Combinat ion 4

Figure 57. Combinations of weights applied to Group 3

Group4

2.5

1.5
Error

Land Yacht
Buddhist Monk

0.5

0
Best Relavance+Topology Combination 2 Combination 3 Combination 4

Figure 58. Combinations of weights applied to Group 4

198 Experiments

our model is capable of simulating and making predictions in the context

of established examples. Furthermore, given that, as far as we know, there
are no operational models for the study of blends with a stable, commonly
used, methodology across researchers, Divago could be a starting point for
the validation and discussion of the subset of blends that it can consider. We
believe that such a study would bring some interesting outcomes: (possibly
new) categorizations of blends; understanding of the underlying frames that
are recurrent in some blends; validation of observations previously made (e.g.
Topology is important in the “Buddhist Monk”); predictions regarding novel
blends (e.g. its underlying frames).
It is a fact that, with the knowledge representation used, Divago was able
to find the targets (or very similar solutions) also found in the literature.
Achieving such, it is our opinion that it reached a capacity for making Con-
ceptual Blending, although still at a relatively basic level in comparison to
our own cognition and to the world of examples discussed in (Fauconnier
and Turner 2002), some considering many input spaces, many consecutive
blends, the majority of them not formally described or represented. Clearly,
a more dynamic knowledge representation, perhaps not entirely symbolic,
would be needed to cope with more elaborate, and more realistic, examples
at the level of cognition.

6. Discussion

The experiments presented in this chapter raised a set of questions we would

now like to discuss. We will start by a set of practical issues and progress to-
wards more philosophical questions. We intend to focus on the problems and
virtues of the implementation and of the model, and lead the reader towards
possible evolutions and applications.
From the results obtained (namely regarding nov and use), it is an unde-
niable fact that a few predictions can now be made regarding the behavior
of Divago. We can now say that, with a problem that can be specifiable via
a query, it is able to retrieve a good solution, if this exists in the space of
concept combinations and if some prominence is given to Relevance and In-
tegration weights. We can also predict that Divago will diverge from the in-
puts if this query so implies, or when there are frames available that so imply.
Ultimately, if we have no query or frames available, Divago will have the ten-
dency not to reinvent the inputs, but to generate blends that inherit parts from
 Discussion 199

Table 29. Ritchie’s (Ritchie 2001) measures: summary of Divago + WASP results
Experiment
Measure Horse-Bird Noun-Noun Creatures WASP
1 0.443 0.543 0.343 0.71
2 0.273 0.563 0.333 0.54
3 0.504 0.782 1.000 0.47
4 0.636 0.781 1.000 0.24
5 1.000 0.778 1.000 0.36
6 0.364 0.344 0.667 0.05
7 0.500 0.786 1.000 0.12
8 1.333 0.786 2.000 0.28
9 0.000 0.036 0.000 0.000
10 N/A 16.000 N/A N/A
11 0.406 0.513 0.308 0.71
12 0.483 0.831 1.000 0.47
13 0.273 0.500 0.333 0.54
14 0.636 0.781 1.000 0.24

both, although without any specific overall coherence, namely because of its
space complexity. This raises the issue that, by themselves, the Optimality
Constraints, as we modelled them, did not have success in Divago, except for
Relevance and Integration. In this context, and perhaps in general, we believe
that the eight constraints of Fauconnier and Turner can be reduced to three:
Relevance, for purposefulness; Integration, for internal coherence; and Topol-
ogy/Unpacking, for external coherence. We must acknowledge, though, that
we did not present a thorough account for the Vital Relations and their com-
pression role, which may imply that those three constraints are incomplete
regarding an implementation of CB.
With respect to a direct comparison with WASP (the poetry generation
referred to in section 2.2.2 (Gervás 2001; Gervás 2002)) regarding the values
of Ritchie’s measures, as we said earlier, such an exercise is merely academic
in the sense that, in practice, these are very different from each other. Never-
theless, some new conjectures can be made. First of all, we must provide the
values obtained for WASP, as well as a summary of Divago’s results (figure
29).
200 Experiments

From an analysis of these results, the first thing to conclude is that WASP
clearly produces a higher average of typicality and lower average of value
than Divago (measures 1 to 4). This imbalance also affects measures 5 to 8,
which basically reassures that outcomes of Divago were classified as higher
valued than WASP’s, and measures 11 to 14, which get the same conclusions
by comparison with the inspiring set. Notice that, by measures 9 and 10, nei-
ther system tends to reinvent the inspiring set in any way (i.e. these latter
measures cannot add any new conclusions). We must insist that the actual
values should not be taken further in this comparison. At this point, the most
one can do (and the actual importance of Ritchie’s measures in this case) is
to conclude that, according to the measures used, Divago seems to be more
inventive than WASP. To go further, one would have to compare the specific
evaluation procedures of each system. This would imply a comparison of a
Poetry generation system evaluation methodology (which was based on peo-
ple’s interviews and stylistic analysis) with the ones used in our system. As
this is an unsafe comparison to make, we trust that the evaluations just made
will be more important for future related works (e.g. of creativity assessment
in concept invention systems) than for a competent comparison of Divago
and WASP (to read more about this subject, also including a comparison with
Dupond, read (Pereira, Mendes, Gervás and Cardoso 2005)).
It can also be said that our measures of typicality and value are simplistic
and therefore lead to a high variability in Ritchie’s measures as well as to
some counterintuitive results. As a formal system, Divago needs a set of well
defined criteria and the question is whether the effort of building more com-
plex formulae or heuristics would be justified by an added value in results.
As we said earlier, there can be no universal measure of value, and therefore
following the Occam’s razor principle seems the adequate choice. Moreover,
this choice becomes a virtue of Divago in the sense that this system suggests
a validation that can be applied to different domains, not being tailor-made
for the specific application, as happens in the majority of systems referred to
in this book.
The knowledge representation showed itself to be problematic in some ex-
periments. Namely for the noun-noun combinations and for the established
blending examples, it is clear that our common sense knowledge of the con-
cepts goes far beyond the representations considered, this being one of the
main reasons for such counterintuitive results. Although the frames seemed
extremely powerful, they can never compensate for the poor quality of the
concept maps. Ideally, the concept maps should be dynamic (such as in the
 Discussion 201

Slipnet of Copycat (Hofstadter and Mitchell 1988)) and not isolated. Actually,
this leads to the first strong self criticism we must make regarding Divago.
While, throughout this book we have been arguing for a multi-domain en-
vironment, Divago, as it is implemented, only considers this on a superficial
level. We can see this at two levels. At the level of the individual experi-
ments, the pairs of input concepts considered are rarely distant enough to
each other such that one can unquestionably consider them from different
domains. Still at the level of individual experiments, Divago is given exter-
nally (or randomly) a pair of concepts, and thus it does not “wander” in the
multi-domain environment, but in the space defined by the Blender, which is
much more restrictive. At the level of the overall experiments, it did not con-
sider input concepts from two different experiments (e.g. blending a house
and a horse, or a werewolf with a paper). After verifying the complexity we
faced in the experiments presented, it becomes obvious that our choice for a
set of isolated experiments, some with familiar concepts (horse-bird, house-
boat, creatures), some with less familiar concepts (noun-noun combinations,
established blending examples), comes from a need to observe the capacity
of bisociation of Divago, while avoiding being distracted by other, yet also
important, aspects. We trust that, first of all, in order to reason in a genuine
multi-domain environment, such a system must be able to deal with simpler
situations, with the motivation of gradually being open to a wider scope, as
we made with Divago.
We have been recurrently criticizing the structure alignment algorithm
used in Mapper, but we must add that this is very much unexplored ground
and we have so far found no promising alternatives. The algorithm demon-
strated the virtue of being computationally inexpensive and of proposing
mappings for Divago in some of the experiments.
Another issue to discuss is the interpretation of the blends. We proposed
visual and textual interpretations, yet these describe only a selection of as-
pects, leaving out some others of potential importance. This is rather a prob-
lematic issue. The concept theories and instances were designed to be self
explicatory, however in order to avoid ambiguities, large amounts of knowl-
edge are necessary, and each piece of knowledge recursively demands the
explanation of its constituents, thus demanding the existence of ground sym-
bols. The semantics of these ground symbols must be context dependent (e.g.
the semantics of a wing can be a 3D or 2D image, a functionality descrip-
tion, a word, another network of concepts with f eathers and bones, each one
appropriate in a different context). This means that, while knowledge repre-
202 Experiments

sentation for concepts can be domain independent, their interpretation must

be domain dependent (or at least context dependent).
Divago’s versatility has also been recurrently referred to in that it needs no
structural changes for working with any two different pairs of input concepts.
For each pair of concepts, it demands their description (via concept theory
and instances), a choice (or creation) of frames to use as query and to add to
the generic domain, and a choice of the weights for the Constraints module.
The description of the concepts will always demand some effort, while the
rest may be picked from those available. Of course, for each application, an
interpreter may be necessary, this being the least versatile, although unavoid-
able, aspect.
We have demonstrated also that Divago is divergent in the sense that it
tends not to reproduce the input concepts and agrees with the theories and
principles enunciated in the previous chapters regarding Conceptual Blend-
ing and our Model of Concept Invention. It was empirically demonstrated
that, with an appropriate set of frames (and sufficient Relevance and Inte-
gration) the system is able to produce useful and novel results. Indeed, we
believe frames have an extraordinary power only superficially explored here.
Remember that they may comprise small programs with the expressiveness
of the Prolog language.
In spite of the uncertainties in the assessment of issues like creativity or
divergence, these experiments show that Divago was able to accomplish two
very objective goals: it is able to reach approximately the same results of C3 ,
with a specific set of frames; it can produce the same blends, or approximate
ones, as in the examples listed from the Conceptual Blending literature. The
latter is particularly important as it may become a computational methodol-
ogy for analyzing blends.
As suggested by some of the contexts that we invited the reader to imag-
ine, we assume that the model presented in this book is more useful as a
reasoning mechanism (possibly at the meta-level) that can help a computa-
tional system to extend its space of possibilities, i.e. transform it. Such a
system would need to give our model the description of what a valid possi-
bility involves (via a language such as used in the queries of Divago), which
would then generate bisociations until finding a satisfactory outcome. We ar-
gue that this emulates, at least partially, the process of imagination according
to Koestler, Guilford and Fauconnier and Turner. However, we are aware that
Divago is very incomplete with regard to the implementation of such a model.
As a model of creativity, we have to reassert that it lacks some funda-
 Discussion 203

mental parts, namely the interaction with the environment, which is so fun-
damental, according to Csikszentmihalyi and others. If Divago was a perfect
implementation of the model of Concept Invention discussed here, it would
still be somewhat autistic due to a lack of contact with the external world.
This is another reason why it should not be considered alone and indepen-
dent of a specific purpose or environment.
In conclusion, as far as our definition of creative system given in section
2.2.1 goes, Divago clearly falls into that category. For every experiment made,
it produced more results that are not replications of previous solutions than
copies of its own knowledge; it was able to reach the established goal or just
fall short of it in the majority of the situations. It is based on a cognition-
centered model - the model of Concept Invention, from chapter 18 - and is
implemented as a hybrid AI system, since it applies typical Knowledge Based
Systems techniques (rules, constraints, knowledge representation) as well as
Evolutionary Computation algorithms (the GA of the Factory module). Thus,
one can conclude that Divago is also an AI system, an argument for the thesis
that Computational Creativity should be part of AI, as much as creativity is
part of intelligence.
Chapter 7
Conclusions and Future Directions

Now that we are reaching the end of this book, it is time to draw the main
conclusions, both at the level of Creativity modelling and its many associated
questions that have been referred to since the beginning and at the level of
the practical implementations and models presented here. We cannot finish
without elaborating on the main contributions and pointing to possible future
directions of research.

Modelling Computational Creativity can be seen, by the most skeptical, as

an a priori impossible mission. As a formal machine, a computer is deprived
of aspects that are often considered fundamental, such as intention or emo-
tion. The same argument is also used against Artificial Intelligence. In either
case, one falls into a human-centric view which is monolithic and reduction-
ist at the same time. It is monolithic because it assumes that only a being
(or a thing) that has all the characteristics together in a whole can be consid-
ered creative or intelligent and it is reductionist because it reduces creativity
and intelligence to an all or nothing basis, assuming that only humans fulfill
all of these conditions. From the many studies in the areas of Psychology,
Philosophy or Cognitive Science, some of which have been described here,
we have reached the different conclusion that Creativity is more continuous
than discrete and that it is related to many different aspects, some more com-
putational than others. We have presented arguments for the construction of
Computational Creativity, which does not have to be the same or measured
with the same thresholds as the Human Creativity. Computational Creativity
must be defined more precisely and we have proposed that it should be bound
by the capacity of generating a reasonable amount of valued and atypical so-
lutions to a problem. One conclusion we took from this work was that this
modelling of creativity must not run away from an AI framework. In other
words, one should not avoid facing creativity as a problem of search for so-
lutions to problems (although these become normally much less specified) or
using mechanisms typically from AI (e.g. genetic algorithms, logic program-
ming, neural networks). This does not mean that creativity can only bring
new applications of AI techniques, on the contrary we believe that creativity
is a missing part of AI and, in the same way that humans (and nature in gen-
206 Conclusions and Future Directions

eral) need creativity to be more versatile, less constrained to usual solutions

to usual problems, more open to change and ready to cope with unpredictabil-
ity, the machine will need to use more creative behavioral skills in order to
become more competent.
The model presented here applies many of the well known AI techniques,
such as rule-based systems or genetic algorithms, but these are only the
means to the broader end of modelling bisociation and divergence, which
have not been approached within AI. Traditional AI search has been consid-
ered throughout this work, but always with attention to the world beyond the
search space, and to methods for how to reach it. It is within this paradox of
reaching the unreachable that the study of Creativity can become fundamental
within AI. This book brings some contributions motivated by this quest:
– Model of Concept Invention based on principles and theories from Psy-
chology, Philosophy and Cognitive Science. This model was the leitmotif
of this book, representing an ideal system, as opposed to an actual im-
plementation. It proposes a set of modules and their interaction for the
invention of concepts via the combination of concepts from distinct areas
of the knowledge base. This concept invention is essentially inspired by
Koestler’s bisociation (Koestler 1964) and Guilford’s divergent thinking
(Guilford 1967), while still leaving open other forms of concept creation
and of concept combination.
– Computational model of Conceptual Blending. We present the first ex-
tensive computational approach to Conceptual Blending (Fauconnier and
Turner 1998), which takes into account the several processes (composition,
completion, elaboration, selective projection) and principles (optimality
principles) described in that framework. This model is directly applica-
ble to the study of blends that are based on one-to-one mappings and two
input domains. Once these input domains are represented according to Di-
vago’s representation, they can be tested and analyzed, as we showed. A
specific suggestion we obtained from this computational model of blend-
ing was that the set of optimality principles proposed in the framework are
reducible to a smaller set of principles (Integration, Topology and Rele-
vance).
– Divago. The two models above ultimately led to the implementation of a
system, Divago, which comprises a set of proposals for how to implement
the many modules involved. It has been thoroughly described in this book
and is available for use by other researchers (e.g. for experiments with
Conceptual Blending) or for connecting to another system (e.g. to extend
 Conclusions and Future Directions 207

this other system’s knowledge base). Divago is unique in many aspects,

namely its ability to generate results that are valued (according to a purpose
known to it) yet untypical54 . It is a basic argument of this book that this
tendency to diverge is fundamental for creative behavior.
– Multitude of applications. Divago was tested with a multitude of applica-
tions. If not useful for the applications themselves, for they were more hy-
pothetical situations for testing the system than actually directed to specific
problems, they can be used for comparison later with similar systems and
as a starting point for more specific applications of Divago. More impor-
tantly, they allowed an observation of the system within different situations
and the analysis of its evolution and behavior. They are also proposals for
situations where bisociation can become important and computationally
applicable: the creation of new concepts for another system (a drawing
system in the house-boat; a game system in the creature generation); and
the study of conceptual combination (the noun-noun experiments and the
established blends).
– Creativity assessment. The issue of evaluation is one of the fundamental
problems in the study of creativity. Throughout this book, it has been a pri-
mary concern and, without promising the holy grail of universal formulae
of usefulness or novelty, we propose some ideas for the assessment of cre-
ativity in computational systems, and for systems like Divago in particular.
The main pillars for defining the criteria used in the creativity assessment
of Divago were the works of Ritchie (Ritchie 2001), Wiggins (Wiggins
2001, 2003) and Colton et al (Colton, Pease and Ritchie 2001). We charac-
terized the Model of Concept Invention with the more abstract and generic
perspective of Wiggins, and analyzed Divago with the more concrete per-
spective of Ritchie and Colton et al. The latter led to a precise definition of
typicality, which consists of the distance to the inputs (what is known), and
of value, which is defined by how much the system accomplishes a goal55 .
As these analyses are rare within the field of Creative Systems, we believe
that this work contributes to the evolution of the field in general and to the
problem of assessment in particular.
There is an extensive set of future directions that this research can pursue,
either by us or others and here we list a few that seem to be fundamental:
– Other processes of invention than bisociation. In this book, we have fo-
cused almost exclusively on bisociation as a method for concept invention.
However, other methods may also apply, such as concept re-representation
or interaction with the environment, to name only two candidates. There
208 Conclusions and Future Directions

is no reason to expect that these methods would have to be considerably

different or antagonistic to the one presented here, therefore allowing other
alternatives which could result in the extension of the Model of Concept
Invention with new modules or further exploration of already existing ones.
– Evolutions to the Blending model. The computational model of Blending
presented here should also be subject to further developments, namely the
redesign of the optimality principles, possibly reduced to a smaller set, the
inclusion of the latest changes, namely the vital relations and the compres-
sions. The big leap for this model would be to cope with more than one
input space as well as allow a more realistic knowledge base, which would
have to be extremely large and organized.
– Evolutions to Divago. As this was the main practical part of the work, it
was subject to many compromises, namely paths that we did not follow
for pragmatic and research direction reasons. These paths are deserving of
particular attention in the future:
– Real multi-domain environment. Divago is still not able to work in a real
multi-domain environment since the choice of the pairs of concepts to
bisociate is either made randomly or externally. In a multi-domain envi-
ronment, it should, when facing a problem, make an inspection of all its
knowledge, which would comprise many different domains and knowl-
edge representations, and would be able to pick for itself the sources
for concept invention. Possible algorithms for developing this capability
could come from works on analogy retrieval, where, before establishing
an analogy between two concepts, the system searches for candidates in
the knowledge base (e.g. MAC/FAC, from (Gentner and Forbus, 1995),
or ARCS, from (Holyoak and Thagard, 1997)).
– Meta-level reasoning. Being able to do meta-level reasoning would be a
giant leap for Divago. It would then reason about its own knowledge and
processes, potentially evolving them. A possible inspiration for enabling
meta-level reasoning in Divago could be Simon Colton’s HR (Colton,
Bundy and Walsh 1999), which is able to generate theories about its
theories, and theories about its own rules. Analogously, Divago would
bisociate its own internal rules, such as the optimality constraints or the
blending projections or, a more realistic situation, create new frames
by bisociating existing ones. All this seems extremely complex and de-
manding of a serious research effort.
– Interaction with the environment. It was argued in the beginning of this
book that environment is important for creativity. For some theories (e.g.
 Conclusions and Future Directions 209

the Systems model of Csikszentmihalyi (Csikszentmihalyi 1996)), it is

even a necessary condition for the existence of creativity. In this sense,
Divago is rather autistic and clearly demands more contact with the en-
vironment. Integrating Divago within a multi-agent society environment
seems an interesting project to develop, possibly a hybrid society, as
suggested by (Pazos et al. 2002)).
– Elaboration. The Elaboration phase in concept invention is of great im-
portance. It is there that part of the emergent structure of a new con-
cept is constructed. However, the processes by which concepts are elab-
orated vary a lot depending on the situation. While it is generally agreed
that part of the emergent features come from rule-based elaboration,
i.e. accomplished by straightforward reasoning about a situation (e.g.
a “beach bicycle” must have “large tyres”), other features seem not to
have straightforward explanations (e.g. why does “Dracula” hate gar-
lic?). Alternative processes must be sought for the elaboration, for this
may have a great effect on the creativity of the results. We suggest that
a possible contribution could be the use of other knowledge from the
knowledge base (other than the inputs or the generic domain), for exam-
ple, by searching for similar concepts and bringing new knowledge from
them (e.g. when blending “horse” and “bird”, the result may become
similar to “dragon” and get new knowledge, such as “spitting fire”).
In conclusion, the area of Computational Creativity has been growing in
the past few years and is clearly in its early stages. The current need for
stronger and consensual definitions is notable, as well as evaluation method-
ologies, and benchmarks for comparisons. Its relationship with AI and other
sciences must be established, if it is to gain its own place and flourish to fulfill
its promises. In this sense, our work is only one step in that direction.
Appendix A
The effect of input knowledge

In chapter 6, we make some calculations to estimate the effect of input knowl-

edge in the creativity of Divago. In this appendix, we reproduce the theoreti-
cal basis behind those calculations. Simon Colton, Alison Pease and Graeme
Ritchie (Colton, Pease and Ritchie 2001) propose a set of criteria for evaluat-
ing creativity, now with special attention to the effect of input knowledge.
One of the main problems in evaluating computational creativity (and of
AI systems in general) relates to the extent to which the system’s knowledge
is fine-tuned, i.e. the system essentially replicates known items to a greater
extent than it causes the generation of novel high-valued items (Colton, Pease
and Ritchie 2001).
Let OK be the set of output items corresponding to input knowledge K.
We define VK as the set of high-valued items in OK ; RK are the reinventions
(the items that belong to the inspiring set I); and CK is the creative set (the
items in VK which were not in RK ).
In order to determine the effect of a subset K of K (the input knowledge),
let us first examine the possible effects on V(K−K ) :
– K is creatively irrelevant if VK = V(K−K ) .
– K is creatively useful if V(K−K ) ⊂ VK .
– K is creatively destructive if VK ⊂ V(K−K ) .
For the creatively useful K , Colton et al. define the dependency set DK ,
such that DK = VK −V(K−K ) , which corresponds to the set of items that will
not be obtained if we remove K from input knowledge. We can now say that
K is fine-tuned if (Colton, Pease and Ritchie 2001):

|DK ∩ RK | > 0 and |DK ∩CK | = 0

This means that the presence of K in input knowledge only contributes to the
replication of high-valued items, without having influence in the production
of creative outcomes. For cases where K still contributes to creativity (i.e.
|DK ∩CK | > 0), we can obtain a measure of how fine-tuned K is:

|DK ∩ RK |
f t(K ) =
|DK ∩CK |
212 A. The effect of input knowledge

Naturally, when f t(K ) is greater than 1, it means that K is used to redis-

cover more items than to generate new ones of value. In order to determine
whether a program P is fine-tuned when using knowledge K, Colton et al.
propose the following two measures (assuming P was constructed using in-
spiring set I):
|K | K
– m1 (P, I, K) = |K|f t , where K f t =
K ⊂ K : K is fine-tuned
– m2 (P, I, K) = max( f t(K )) over K ⊂ K
If m1 is greater than 0 or m2 greater than 1, we can claim that P using K
has been fine-tuned to some extent. If m1 is 1, P using K is completely fine-
tuned. If m2 is greater than 1, then there is at least one such subset of K which
is used more to replicate known artifacts than to find new ones (Colton, Pease
and Ritchie 2001).

Some of the measures presented in this section were applied to the work
presented in this book. Maybe due to being quite recent and still demanding
refinements of many sorts, these measures have not been applied in practical
computational systems, with the exception of the analysis of Pablo Gervás to
his poetry generation system, WASP (Gervás 2002), who applied Ritchie’s
measures, and of Colton’s HR fine tuning analysis (Colton, Pease and Ritchie
2001).
Appendix B
Established examples of Conceptual Blending

In this appendix, we will show the established blending examples that are
used in chapter 5. In chapter 3, we have already described two examples
(“Riddle of the Buddhist Monk” and “computer virus”). As with those, we
reproduce the diagrams, tables and explanations as close as possible to the
original ones. The “CEO boxing fight” example has two input spaces with
different organizing frames (Fauconnier and Turner 2002: 128-129). It is a
metaphoric scenario that conceptualizes business competition. According to
this metaphor, we can say that “one CEO has landed a blow but the other one
has recovered”, “one of them knocked the other out cold”, etc. In other words,
it is the structuring of the business domain according to the boxing domain.
In figure 59, we show the corresponding network as proposed by F&T. Since
only one input space determines the organizing frame of the blend, this is a
single scope blend.
The trashcan basketball example (in figure 60) refers to the “game” one
imagines to play when throwing papers at the wastepaper basket (Coulson
2000: 118-119). This involves the integration of the domains of Basketball
(imagination) and trash disposal (reality). The emergent structure arises from
affordances in the environment. In trashcan basketball, some elements are
inherited from trash disposal domains (“trashcan”, “crumpled-paper”) while
others from the basketball domain (“shoot(person, paper, trashcan)”). Since
the structure of the game comes from both domains (the rules of basketball,
the affordances of the room), this is considered a double-scope blend.
The next example is a nominal compound that leads to a permanent cat-
egory change. More precisely, the notion of “same-sex marriage”. One input
space is the traditional scenario of marriage, while the other describes an al-
ternative domestic scenario involving two people of the same sex (Fauconnier
and Turner 2002: 271-272). The cross-space mappings may link typical ele-
ments such as partners, common dwellings, commitment, love, sex. Selective
projection then pulls to the blend social recognition, wedding ceremonies and
mode of taxation from the traditional marriage input, while same sex, absence
of biologically common children and culturally defined roles of the partners
are projected from the other input (see figure 61). Thus, this is also a double-
scope blend.
214 B. Established examples of Conceptual Blending

Generic
Competition Space
between
competitors

boxer 1 CEO1

boxer 2 CEO2
one boxer CEO1 defeats
knocks out the CEO2
other

Boxing
boxing Business
CEO1
boxing
CEO2

CEO1 knocks
out CEO2
Boxing
CEOs
Figure 59. The blending diagram of CEO fight

Another very classic example is known as the “Debate with Kant” (see
figure 62) . It is about the following monologue (more precisely, an imagined
dialogue) (Fauconnier and Turner 2002: 62-63):

I claim that reason is a self-developing capacity. Kant disagrees with me on

this point. He says it’s innate, but I answer that that’s begging the question,
to which he counters, in Critique of Pure Reason, that only innate ideas have
power. But I say to that, What about neuronal group selection? And he gives
no answer.

In one input space, we have the modern philosopher (m) making claims,
aware of Kant (k2 ) - the eighteenth century philosopher. In a separate input
space, we have Kant (k1 ) - the living philosopher -, thinking and writing. The
blended space has both people and the “debate” frame has been recruited
since there is no debate in either input. The debate frame comes to the blend
 Established examples of Conceptual Blending 215

Generic
Space
vaguely spherical (b*)
Bball(b*)

crumpled
leather-
paper (b’’)
sphere(b)
Bball(b) trash(b’’)

shoot(person, throw-out(person,
bball, basket) paper, trashcan)

Basketball
Trash disposal
crumpled-
paper (b’)

trashcan-
Bball (b’)

shoot(person,
bball, basket)

Blend
Figure 60. The blending diagrams of Trashcan Basketball

through pattern completion, since so much of its structure is already in place

as a result of composition (i.e. many of the elements and relations of the
debate frame were already in the blend before it was recruited). Once the
blend is established, we can “run the blend”, in this case, this is done by
instantiating the debate frame with arguments from both input spaces.
In the next examples (“gun wound”, “pet fish” and “land yacht”), Seana
Coulson approaches a subject that is typical of Conceptual Combination:
noun-noun compounds. She thus proposes applying the Conceptual Blending
to explain the (conventional) meanings of each of the compounds. A “gun
wound” is a “wound” (directly or indirectly) caused by a “gun”. The strat-
egy followed to deconstruct this compound is by recruiting the action frames
associated with each domain. The generic space contains the generic Vio-
lentAct frame with a cause and a result, while the blend contains the more
216 B. Established examples of Conceptual Blending

man and woman same-sex couple

religious sacrament
legal status
procreation
children
family
kinship
socially positive status stigmatized
publicly displayed often concealed
love love
sex sex
live together
live together
finantial advantages
...

Same-sex
Traditional partnership
same-sex couple
marriage
religious sacrament
legal status
socially positive status
publicly displayed
love
sex
live together
finantial advantages
... Same-sex
marriage

Figure 61. The blending diagram of Same-sex marriage

specific Shoot frame. The input spaces bring the cause (the “gun” domain)
and the effect (the “wound”) (Coulson 2000: 130-131) (see table 30).
“Pet fish” is a blend in which the two inputs (“Pet” and “Fish”) are coun-
terparts and map onto the same element in the blended space, i.e. they fuse
into the same concept. As in many examples already given, the blend inherits
structure from both input spaces. Knowledge having to do with “pet owner-
ship” are inherited from the “Pet” domain, while fish attributes come from
the “Fish” domain (Coulson 2000: 143-44) (see table 31).
 Established examples of Conceptual Blending 217

thinker
claims and musings

mode of expression

language

issue
purpose

time

Generic
me (m)
Kant [k1]
claims and musings
claims and musings
writing speaking
English
German
cognitve processes
reason
search for truth
search for truth 1995
1784 Kant [k2]
k2 dead
k2 never aware of m2
m2 aware of k2

Input 1 k Input 2
m
claims, counterclaims,
questions, answers,...
speaking Debate FRAME
English (same language)
rhetorical actions:
cognition agrees, retorts,
search for truth challenges,
anticipates…
same time t
argumentation
m, k alive connectives, affirmation
m aware of k and negation:
but, however, therefore,
Blend k aware of m one the contrary...
...

Figure 62. The blending diagram of Kant Debate

218 B. Established examples of Conceptual Blending

Table 30. Gun Wound mappings

Input1 Input2 Blend Generic
Gun Wound GunWound ViolentAct
Elements Elements Elements Elements
Agent DangerousAct Agent ViolentAct
Target Human Human Patient
Gun Means Gun Means
Result Wound GunWound Damage
Relations Relations Relations Relations
Shoot(Agent, Cause(Dangerous Shoot(Agent, Cause(ViolentAct,
Act,
Gun, Target) Means, Human) Gun, Human) Means, Patient)
Result(Result) Result(Wound) Result(Gun Result(Damage)
Wound)

Table 31. Pet Fish mappings

Input1 Input2 Blend
Elements Elements Elements
Pet Fish PetFish
Owner Water Owner
House House
Tank
Relations Relations Relations
Feeds(Owner, Pet) Lives-in(Fish, Water) Feeds(Owner, Fish)
Loves(Owner, Pet)
Swims(Fish) Swims(Fish)
Lives-in(Fish, Tank)

The “Land Yacht” compound demands more subtle reasoning. A land

yacht is a very high class luxury car, it inherits the central properties of car
and the diagnostic properties of yacht (in relation to other sailboats, a yacht
is a luxury boat, extremely expensive and providing high social status to the
owner). Therefore, the projection from inputs is more unbalanced than was
the case with “Pet Fish”, where the central properties of each were projected
to the blend. Here, incongruities would arise if so happened (e.g. a car can-
 Established examples of Conceptual Blending 219

Table 32. Land Yacht mappings

Input1 Input2 Blend
Land Yacht Land Yacht
Elements Elements Elements
land water land
driver skipper driver
road course highway
car yacht luxury car
owner tycoon rich owner
Relations Relations Relations
Drives(driver, car, Sails(skipper, yacht, Drives(driver, car,
road) course) highway)
Yacht Luxury Car
Function: sails Function: drives
Sign-of: upper-class Sign-of: wealth
Owner: tycoon Owner: rich person

not sail). Table 32 shows the cross-space mappings as Coulson presents them
(Coulson 2000: 155-156).
The example of the “computer desktop” comes from Tim Rohrer, who
is interested in the relationship of Metaphor with information technologies
(Rohrer 2000). The computer desktop interface comes as a metaphorical pro-
jection of a physical desktop in an office, with folders, storages, waste basket,
documents and the respective actions (moving physical objects from different
places, opening folders), to computer data management representation (direc-
tories, files) and physical objects (screen, drive). Table 33 is, therefore, not
descriptive of the blending process. It basically shows the direct correspon-
dences between the desktop with the blend, leaving implicit the computer
domain elements.
Counterfactuals are also a recursive theme in Blending literature (Lee and
Barnden 2001)). Counterfactuals are statements about the consequences of
things that happen to be false (e.g. “If I were you...”). We present one of
Seana Coulson’s counterfactual examples: “If I had bread, I could make a
sandwich”. The inputs are the Actual space (Seana has turkey, cheese and
mustard in the fridge), and a Sandwich space, in which there is bread, condi-
ments, meat and cheese (Coulson 2000: 106-107) (see table 34).
220 B. Established examples of Conceptual Blending

Table 33. Computer Desktop mappings

D ESKTOP H UMAN -C OMPUTER I NTERFACE
(input 1 - source domain) (blend - target domain)
Desktop → Screen
Documents → Files
Folders → Directories
Storage → Drive icons
Moving physical objects → Dragging icons
Putting physical objects down → Dropping icons
Deleting objects → Dropping icons in recycle bin
Focusing on a task → ’Zooming in’, opening window
Putting away a task → ’Zooming out’, closing window

Table 34. Sandwich counterfactual mappings

Input1 Input2 Blend
Actual Sandwich Counterfactual
Seana Agent Seana’
Fridge Fridge” Fridge’
Turkey Turkey” Turkey’
Cheese Cheese” Cheese’
Mustard Mustard” Mustard’
Bread” Bread’

According to the author, this is a single-scope blend because the organiz-

ing structure of the counterfactual comes from the Sandwich space, i.e., it
states the individual roles of each element.
The final example we show is also an established theme in blending lit-
erature: the “X is the Y of Z” constructions. In this case, it is instantiated as
“Mabel is the daughter of Paul”(Coulson 2000: 119-120). It is also a single
scope blend, since structure comes only from input 2 (see table 35).
 Established examples of Conceptual Blending 221

Table 35. “Mabel is the daughter of Paul” mappings

Input1 Input2 Blend
Elements Elements Elements
Mabel Daughter Mabel
Paul Father Paul
Relations Relations
Daughter-of(Daughter, Daughter-of(Mabel,
Father) Paul)
Appendix C
Programming the Frames

1. Syntax and Overview

The syntax of every frame is:

Predicate: frame(Domain, Name, PosConds, NegConds, PosConc,
NegConc)
Domain identifier of the domain or concept in which
the frame is included. (e.g. generic, eating)
Name name of the frame (e.g. aprojection, aframe)
PosConds Positive premisses (e.g. have(X, wings))
NegConds Negative premisses (e.g. weight(X, very heavy))
PosConc Positive conclusions (e.g. ability(X, fly))
NegConc Negative conclusions (e.g. habitat(X, water))
In first order logic, the frame can be represented as:
PosConc and not (NegConc) ←− PosConds and not(NegConds)
For example (prolog like):
ability(X, fly), not (habitat(X, water)) :- have(X, wings), not(weight(X,
very heavy)).
Applied to the blend, this would mean that, if X has wings and it is not very
heavy, then it has the ability to fly and its habitat should not be water. If we used the
above frame as a query to the blend (measured within the “relevance” optimality con-
straint), it would tend to evolve towards a blend that has the element wings projected
(and so the relation has( ,wings) as a consequence) while the very heavy element
is not projected (it can be projected, but it must be in such a way that weight( ,
very heavy) is not)56.
Later, in the “elaboration phase”, the conclusions are triggered, so the predicate
ability(X, fly) is added, while the predicate habitat(X,water) is removed (if it exists).

2. Programming of the frames

Four kinds of terms are allowed in any of the frame conditions (or conclusions) part:

1. Regular concept map binary predicates, like “sound(X, neigh)” or “purpose(Y,

fly)”. There are also special variations in which we can give multiple op-
tions for a predicate parameter, specified inside a list. For example, “sound(X,
[neigh, bark, chirp])” means that if X neighs, barks or chirps, then the con-
dition is satisfied. A specific binary predicate, isaN, is also considered, being
224 C. Programming the Frames

the transitive closure for isa (e.g. “isaN(human, animal)” because “isa(human,
primate)” and “isa(primate, mammal)” and “isa(mammal, animal)”).

2. Projection specifications, with the syntax projection(Concept, Origin, Desti-

nation). And the meaning of this condition is that element Origin, which be-
longs to the concept Concept, should be projected into element Destination,
in the blend. For example, if we had the condition “projection(horse, neigh,
neigh)” in a frame, it means we are requiring the element “neigh” to be kept
in the blend as it is in the (input space) horse concept. Of course, this doesn’t
mean the neighbour elements also keep their original “names” (e.g. it is pos-
sible to find “produce(beak, neigh)”, instead of “produce(mouth, neigh)”.

3. Special operators, with the syntax op(Operator), where Operator corresponds

to any command the frame interpreter should understand. Currently, we
have only one single operator: exists(List). This operator converts the ele-
ments of List into the binary predicates and projection format. For example,
op(exists([sound/X/neigh, purpose/Y/fly, projection/horse/neigh/neigh])) will
just convert the list into “sound(X, neigh), purpose(X, fly), etc” and add it to
the list of conditions of the frame.

4. Prolog calls, inside curly brackets (“{}”), just as in DCG grammars syntax.
This allows the programming of frames in “regular” prolog. Naturally, some
specific predicates have been created, for situations that happen regularly in
frame programming:
– stats(D,X) yields some statistics of the current blending operation (e.g.
stats(domain1, X) returns the identifier of domain 1; stats(frame, f)
means that the frame f is satisfied in the blend)
– current blend(Blend) Blend is the identifier of the blend being created
– m(R, X, Y) Returns the mapping correspondences according to the vital
relation R (e.g. m(analogy, horse, bird) means there is mapping between
horse and bird, according to analogy)
– rel(D, X, R, Y) Direct access to the concept map of domain D (e.g.
rel(horse, legs, quantity, 4))
– projection(B,C,X,Y) Direct access to the projection predicates, where
B is the blend in which X, from concept C is projected into Y.
– other input domain(D1,D2) Given D1 or D2, instantiates the other with
the “opposite” input domain (e.g. other input domain(bird,X) ’ X=
horse)
– relationArc(Concept, Action) True if Action has a action/actee
configuration in Concept (e.g. relationArc(eating, eating), relation-
Arc(basketball, shoot)”
– descendant(R, GUMConcept) R is descendant of GUMConcept in the
GUM hierarchy (e.g. descendant(snowing, ambient process))
 Examples 225

3. Examples

We show some example frames, organized according to the level of abstraction. Sim-
ple frames only use “regular” binary predicates, while intermediate frames already
apply predicates in brackets, still connecting to “lower level” reasonings. The ab-
stract frames deal with the reasoning behind the blend construction (e.g. “project
elements from one domain, while maintaining the structure of the other”).

Simple frames
Name:haunted
Code: frame(generic, haunted,
[contain(X, Y), cause effect(Y, fear), attribute(X, [magic, unknown])],
[],
[property(haunted,X), cause effect(haunted, interesting)],
[]).
Description: Something is haunted if it contains something that causes
fear and is magic or mysterious
Name: artefact
Code: frame(generic, artefact(X),
[isaN(X, physical object), purpose(X, Y), isaN(Y, task)],
[],[],[]).
Description: X is an artefact if it is a physical object whose purpose is a
specific task
Name: habitat earth
Code: frame(generic, habitat earth(X),
[place(X, [land, earth, ground, solid])],
[],[],[]).
Description: The habitat of X is earth if its place is in either land, earth,
ground or solid
Name:habitat water
Code:frame(generic, habitat water(X),
[place(X, [sea, ocean, water, liquid])],
[], [], []).
Description:The habitat of X is water if its place is in either sea, ocean,
water or liquid

Intermediate frames
Name: amphibious
226 C. Programming the Frames

Code: frame(generic, amphibious(X), [stats(frame,

habitat water(X)), stats(frame, habitat earth(X))],
[],
[isa(X, amphibious)],[]).

Description: X is an amphibious if it satisfies frames habitat water(X) and

habitat earth(X)
Name: new ability
Code: frame(generic, new ability(D1),
[ability(X,A), purpose(P, A), pw(P,X),
{current blend(Blend), projection(Blend, D1, X, X),
other input domain(D1,D2), projection(Blend, D2, A, A)},
op(exists([projection/D1/X/X, projection/D2/A/A]))],
[{rel(D1, X, ability, A)}],
[new ability(X, A)], []).
Description: An element projected from concept D1 has a “new ability”
in the blend. I.e. there is a X (projected from D1) that has an ability A that
didn’t exist in D1 and was projected from D2. In order to be a “well founded”
ability, X must have a subpart (P) that enables it to do A.
Name: quality transfer
Code: frame(generic, quality transfer(D1,Q),
[{current blend(Blend), rel(Blend, X, Q,A), descendant(Q, simple quality)},
{projection(Blend, D1, X, X), other input domain(D1,D2),
projection(Blend, D2, A, A)},
op(exists([projection/D1/X/X, projection/D2/A/A]))],
[{rel(D1, X, Q, A)}],
[new quality(X, A)], []).
Description: There is a quality transferred from concept D2 onto D1. In
the blend, there is an element X that has a relational quality (i.e. a relation
that descends from the simple quality node in GUM hierarchy) A that didn’t
exist in the original space.
Name: living thing personificationA
Code: frame(generic, living thing personificationA,
[actor( ,A), {current blend(Blend), projection(Blend, D, A, A), rel(D, A, isa,
Type)}],
[isaN(Type, living entity)],
[personification(A, living thing)],[]).
 Examples 227

Description: A is able to be an actor of some action, and this becomes a

personification of a living thing if A is not a living thing (e.g. a “actor(eating,
pencil)” - we are doing a personification of the pencil, which is not a living
thing)
Name: living thing personificationB
Code: frame(generic, living thing personificationB,
[ability(A, ), {current blend(Blend), projection(Blend, D, A, A), rel(D, A, isa,
Type)}],
[isaN(Type, living entity)],
[personification(A, living thing)],[]).
Description: The same as above, but A is not expected to be an actor, but
instead it should have an ability (e.g. “ability(book, fly)”)

Abstract frames
Name: aprojection
Code:frame(generic, aprojection(A),
[{stats(domain1, A), current blend(Blend),
findall(projection/A/X/X, (projection(Blend,A,X, )), L1)},op(exists(L1))],
[],
[aprojection(A,Blend)],[]).
Description: Every element from domain/concept 1 (A) should be pro-
jected (unchanged) to the blend. For example, in “aprojection(horse)”, every
single element of “horse” (legs, mouth, snout, mane, neigh, run, cargo, pet,
etc.) should be present in the blend.
Name: analogy transfer
Code: frame(generic, analogy transfer,
[{stats(domain1, A), stats(domain2, B),
findall(projection/A/X/Y, (m( , X, Y), not(relationArc(A, X))), L1)},
op(exists(L1)),
{findall(projection/B/Y/Y, (m( , , Y), not(relationArc(B, Y))), L2)},
op(exists(L2))],
[],
[analogy transfer(B,A)],[]).
Description: Every mapped element X that are part of concept A should
be projected to their counterpart Y of concept B (except when X corresponds
to an actor/actee action name, e.g. “eating” is projected to “eating” and not to
“reading”). And every element Y from concept B should also be projected to
Y in the blend.
228 C. Programming the Frames

Name: aframe
Code: frame(generic, aframe(A),
[{stats(domain1, A), current blend(Blend),
findall(R/X/Y, (rel(A, XA, R, YA), projection(Blend, A, XA, X),
projection(Blend, A, YA, Y)),L1)}, op(exists(L1)),
{findall(projection/A/Action/Action, relationArc(A, Action), L2)},
op(exists(L2))], [],

Description: Every relation R that is present in domain/concept 1 (A),

should also be present in the blend, regardless of the projection of the argu-
ment elements (e.g. the ability relation in “ability(bird, fly)” should be present
in the blend as in “ability(horse, fly)”). Once again, there is the special case
of actor/actee relation descriptions, which should also be projected (e.g. if us-
ing aframe in blending “basketball and trash disposal”, “shooting” should be
projected to the blend, as well as the relations from the basketball domain).
Name: noun noun
Code: frame(generic, noun noun(A,B),
[{stats(domain1, A), stats(domain2, B), L=[projection/A/A/A,
projection/B/B/B]}, op(exists(L)), isa(B, Something), projection(A,C,B),
{C\=B}, {(rel(generic, C, isa, Something); rel(A, C, isa, Something))}],
[], [], []).
Description: A proposal of a noun-noun combination. The idea is that A
and B have a relational connection such as “property of”, “lives in”, etc.
For the example “house dog”, where A=house and B=dog, the frame im-
plies that “house” and “dog” are projected to the blend exactly as they are
(they remain the same elements), but there is also a projection from the
“house” domain to the “dog” (e.g. “projection(house, person, dog)”), which
yields a dog that inhabits a house ’ which is one of the interpretations for
“house dog”
In Figure 63, we give an idea of the application of frames to a blend (to
improve readability, both the frames and the elements are simplified). We say
that the blend accomplishes (or satisfies) “aframe”, “transport means” and
“new ability” and that its overall frame coverage is 100% (every relation is
included in a frame). The coverage of “aframe” is aproximately 72% (the
ratio of the blend that is covered by “aframe”) and “transport means” and
“new ability” have coverages of 54% and 27% (respectively).
 Examples 229

Figure 63. Three frames in a small blend

Appendix D
Instances in the house-boat experiment

In this appendix, we describe the language used in the instances for the house-
boat experiment.
An instance follows a hierarchical case representation, each node in the
hierarchy written in prolog-like form:
case(Instance Name, Node Address, Node Name,Cmd List).
with Instance Name being the identifier of the instance; Node Address a
unique identifier of the node within the instance with respect to its position;
Node Name an identifier of the node within the instance (which can be re-
peated); and Cmd List the list of commands that correspond to the semantics
of the specific node. Therefore, this representation is structured top to bottom
(the attribute “son” indicates the descendants of a node):
– each level adds a number to the address (e.g., 0 is the root node, 0:0 is the
first son of the root node, 0:0:1 is the second son of 0:0)
– each node in the structure corresponds to an area of the drawing, containing
one or more shapes.
– some shapes are pre-defined (e.g., parallelogram boat, oval, rectangle, etc.)
in Logo.
– each shape position is relative to a reference point (“in” indicates the com-
mands to apply from the reference point to the starting point of the shape),
normally the upper right corner of the smallest rectangle that can include
the shape
The Logo keywords for defining shapes are:
– left/X. Rotate 45 degrees left
– right/X. Rotate 45 degrees right
– on/X. Move and write X pixels in the current direction
– off/X. Move without writing X pixels in the current direction
For example, the function that defines the shape triangle(X ) is defined
by the list [on/X , right/120, on/X , right/120, on/X , right/120]. The used
shapes are all defined in the file “logoCommands.pl”. The representation for
sailing boat goes like this:
case(b1,0,sailing boat,[sons=3,size=small, type=simple,son name=vessel,
son name=mast,son name=sail]).
case(b1,0:0, vessel, [sons=2, in=[left/90,off/14,right/90],
232 D. Instances in the house-boat experiment

son name=floating structure, son name=hatch]).

case(b1,0:0:0, floating structure, [shape=parallelogram boat, size=
small]).
case(b1,0:0:1, hatch, [shape=oval(5,5), size=small, in=[off/25,right/90,
off/6,left/90]]).
case(b1,0:1, mast, [shape=rectangle(4,30), type=very thin, in=[off/18]]).
case(b1,0:2, sail, [shape=triangle(30), in=[off/18, right/90, off/7,
right/90, off/6, off/13, right/180]]).

And the house is represented as:

case(1,0, house, [sons=2, size=small, type=simple, son name=roof,

son name=body]).
case(1,0:0, roof, [shape=triangle(30)]).
case(1,0:1, body, [sons=3, in=[left/90,off/25, right/90],son name=
structure, son name=window, son name=door]).
case(1,0:1:0, structure, [shape=square]).
case(1,0:1:1, window, [shape=square(5), in=[off/20, right/90, off/15,
left/90]]).
case(1,0:1:2, door, [shape=rectangle(4, 10), in=[off/3]]).
Appendix E
Experiments, Databases and other documents

This appendix corresponds to the CD that comes attached to this document.

It has the directory structure shown in the figure 64. The reader will also find
“readme” files with explanations about some of the files and directories.

Figure 64. The directory structure of the CD.

234 E. Experiments, Databases and other documents

At the website of the book (http://eden.dei.uc.pt/∼camara/AICreativity),

the reader can also find the most updated versions of related software and
complementary publications.
Notes

1. Technically, the solution space can be seen simply as the set of solutions to the problem.
The search space corresponds to the ordered set of solutions. This ordering can be given
by heuristics, distance to the initial situation, or any other criterion that can drive the
search.
2. Particularly in multi-agent systems approaches.
3. As cited in (Martindale 1999)
4. As cited in (Albert and Runco 1999)
5. This example will be further explained in the next section.
6. There are countless examples in which the solution appeared but the scientist wasn’t able
to understand it or to see it as such.
7. Csikszentmihalyi’s systems model also falls into this category
8. Here we emphasize the originality aspect of divergent thinking, as it was defined by
Guilford, giving flexibility, fluency and elaboration a secondary role. In our opinion,
flexibility would solely depend on originality, for we can only get varied solutions if
each one is sufficiently different from the others, i.e. original. On the other hand, we
see fluency as a characteristic of the thinker, not of the thought itself (a thought can be
original, but never fluent). Finally, analyzing past work, elaboration has been considered
belonging to the convergent side.
9. This definition of logical reasoning comes from a psychology perspective, therefore it
may be incomplete from an AI logician point of view. However, we cannot describe
these concepts more formally than allowed by the literature itself.
10. We remind that Csikszentmihalyi allows a kind of personal creativity, which is though
secondary for the system and for what he calls true creativity.
11. Although Ritchie has named them “criteria”, in practice they do not behave as such: none
is an explicit condition for (or even against) creativity. In order to avoid these criteria to
be seen as desired characteristics of creative artifacts, we simply name them as measures.
The assumption is that they are simply numerical estimates that can help characterize the
creativity of a system.
12. He proposes t-creativity as the situation described by measure 6 (high val, low typ), but
for which there is a (yet to be defined) rating scheme typ that highly classifies the items.
13. We conventionalize “recently” to comprehend the past four years, for it is since 2000 that
events exclusively dedicated to the subject have been held yearly.
14. These two cases are more specifically interactive genetic programming (IGP) systems
15. For example, the use of triangular number (from number theory) to predict the order of
the duplicated node (in graph theory).
16. It should be said that these systems are many times built with an analytical intention (e.g.
for prediction or classification) rather than for generative purposes, although becoming
able to be used as both.
17. Thus the name “conceptual invention”, given in some contexts (e.g. the Learning Cycle
of Lawson-Abraham (Lawson, Abraham and Renner 1989) contains a step in which “the
students and/or teacher derive the concept from the data, usually a classroom discus-
sion (the CONCEPTUAL INVENTION phase)”) seems now unfortunate. Words such as
236 Notes

“discovery” or “formation” are clearly less ambiguous.

18. In this context, a concept is identified by a node, but its definition comes from its rela-
tionships with the other concepts, which is consistent with the Theory view given at the
beginning of this chapter.
19. As cited in (Chandler 2002)
20. As the reader will see in the conclusions, we also consider the reduction of the number
of principles to three: Relevance, Topology and Integration.
21. Of course, in (Fauconnier and Turner 2002), this constraint is more explained, but from
the beginning to the end we keep confused with the lack of clear difference to Maximiza-
tion of Vital Relations.
22. A fact or a theory being falsifiable means that it can, in principle, be proven false. Ac-
cording to Karl Popper (Popper 1959), a theory that explains a phenomenon must be
falsifiable through an experimental result that is implied by the theory. One can justify a
theory rationally if it is not (yet) falsified.
23. A research program consists of a theoretical core which is protected by a belt of auxiliary
hypotheses. When the theory is falsified, an auxiliary hypothesis should be reconsidered,
not the theoretical hard core. A revolution is explained as a change in the theoretical hard
core (Lakatos 1978). Lakatos elucidated the activity of science not as the project of trying
to refute one theory but as investigating empirical phenomena within the theoretical frame
of a research program (in (Bosch 2001)).
24. Another common criticism is that the authors of Conceptual Blending have not given
fair credit to preceding works, such as James Hampton’s Composite Prototype Model
(Hampton 1987), which, as can now be understood, stated the same general lines fol-
lowed ten years later by Conceptual Blending, although it is also fair to acknowledge
that Fauconnier and Turner have made much deeper explorations into the same issues.
25. This claim is nowadays widely agreed, as metaphor is seen as a cognitive rather than
a linguistic device. For an extensive figurative versus literalist analysis, we redirect the
reader to (Veale 1995)
26. It is important to give reference here to the works of P. Winston (Winston 1980) and T.
Evans (Evans 1968) which laid the foundations for the study of Analogy in AI, particulary
the Structure Mapping approaches.
27. I.e. if we have MH(A,B) and a relation r, with (r A C) and (r B D) then the collection
must also contain MH(C,D)
28. There is also the incremental rule, which is only applied in successive runs of Sapper,
taking advantage of previous traces, but this is only a lateral subject here.
29. The system pretends that a given interpretation, however ridiculous (e.g. John having
literally one person in each part of the mind), can be real. In doing so, the system is said
to be “semantically agnostic”.
30. The reader may have noticed that we have a clash here: in section 2, we associated R to
Ritchie’s typ and E to val. This would imply the correspondences use↔typ and nov↔val.
An alternative would be to map R to nov and E to use. This is incoherent and demon-
strates the ambiguity in the meaning of R and E . The problem seems to be that E should
not be independent of R, as much as it shouldn’t be independent of purpose (or useful-
ness).
31. We could consider UKB as the inspiring set, since it is given a priori. However, applying
an element from the inspiring set (i.e. a known concept) in an untypical way (e.g. using
 Notes 237

an apple as a weapon, instead of food) would thus be a reinvention by some measures of

Ritchie. Moreover, it is common to have elements in the inspiring set that are not part of
UKB , as well as conversely.
32. Of course, this depends on the mapping functions and transfer operations used, but we
are considering any kind of functions or operations, even randomness, which would be
likely to produce many inconsistencies.
33. This classification is being used for this book, as the use of the names “domain”, “con-
cept” and “concept map” has raised some ambiguity in preceding publications. We will
follow the present definition throughout this document and future publications. It is also
important to inform that the notions and underlying rationales maintain the same.
34. The name element is also used by Fauconnier and Turner to refer to the same entities
inside mental spaces.
35. Rules can also contain Prolog scripts. While in frames, we use complex scripts (e.g.
for specifying what kind of projections to apply in a blend), for rules, we normally use
simpler ones (e.g. for calculating the meeting point of two moving objects, as needed in
the Buddhist monk example.).
36. They are not mandatory, their role is to identify logically incongruous or inconsistent sit-
uations. But if there is a good reason for their maintenance, they will prevail (as happens
in some creativity situations, as mentioned in section 1.2).
37. Assuming n as the number of elements of the largest (in number of elements) of the two
concept maps, we will have a search space of at most n! possible mappings. So, with
n
an exponential kn , as n approaches infinity, kn! will approach 0, meaning that the search
space will expand more than exponentially as the number of elements grows. In little-o
notation, we have that kn is o(n!).
38. Actually, the latest version of Sapper, which has no “prime factorization or wave” em-
ploys only one phase (Veale and O’Donoghue 2000). But we based ourselves in Veale’s
PhD thesis work, thus we will focus on that version. Furthermore, as said before, the
Sapper project has much more testing and documentation for the first version than for the
latter one.
39. A mapping of size i will have i correspondences between input concept 1 and input
concept 2.
40. This helps to understand why, in spite of this combinatorics analysis, the actual number
of different mappings generated is rather small (e.g. for the concept maps in tables 7 and
6, the Mapper generates only three mappings, see fig. 9).
41. A first approach was published in (Pereira and Cardoso 2003b), followed by (Pereira
and Cardoso, 2003c). The current version is a revision with differences in Topologys,
Unpackings and Intensification of Vital Relations
42. To know more about compression, please read (Turner 2006).
43. In fact, an attempt was made in (Pereira and Cardoso, 2003c; Pereira 2005) to model
Intensification of Vital Relations, yet this proposal, which assumes more than one type
of mapping, could not be rigorously tested or further explored.
44. Change, identity, time, space, cause-effect, part-whole, representation, role, analogy, dis-
analogy, property, similarity, category, intentionality and uniqueness.
45. We are calling body to the four walls that hold the house
46. In the experiments reported in this book, we applied the revised version of the constraints,
therefore the results here may slightly differ from the ones given in the paper.
238 Notes

47. These rather large values for mutation and randomness result from the fact that some
mutations have a null effect in the projection scheme (e.g. if an element projection is
mutated to nil, it won’t have effects if its surrounding elements are already projected to
nil). In order to stimulate diversity, these values seemed appropriate. Future experiments
are expected to overcome these issues and apply more common values.
48. A run is an entire evolutive cycle, from the initial population to the population in which
the algorithm stops
49. Note: In configuration 2, there is more than one highly scored blend, none with use higher
than 0.59.
50. The vital relations chosen were isa, pw, pur pose and quantity
51. An aspect to refer is that the combinations, as modelled in C3 , are specific to a set of
human languages (English, German, Dutch...). Others, like Portuguese and French, are
less ambiguous because of the obligatory use of prepositions.
52. In this step, we had to extract this mapping ourselves from the 42 combinations, which
was not difficult since every creature has a relatively small set of relations and concepts.
53. Of course, this does not imply that these creatures will never be generated. The system is
only told that such constructions are to be avoided.
54. This being, of course, much depending on the situation, configuration and knowledge
available.
55. The notion of goal here does not imply a thorough definition, for it can be only partially
defined. For example, a thoroughly defined goal can be “draw me a white house, with
two windows, a door and a roof”, while a less defined can be “draw me a construction
where one can live in”.
56. Notice that, in the condition side we consider “negation by failure”, i.e. something is
not true whenever one cannot prove its truth. For example, if there is not the predicate
weight( f ly, very heavy), the we may assume its falsity. On the other hand, on the con-
clusion side, negations imply deletion, i.e. something will cease to be true if the rule is
triggered.
Bibliography

Abelsson, Hal and Andrea diSessa

1981 Turtle Geometry. Cambridge, Massachussets: MIT Press.
Albert, Robert and Mark Runco
1999 A History of research on Creativity. In Handbook of Creativity, Robert Stern-
berg (ed.), 16–35. Cambridge: Cambridge University Press.
Amabile, Teresa
1983 The Social Psychology of Creativity. New York: Springer.
Andersen, Carl Floyd
1996 A Computational model of complex Concept Composition. MSc dissertation,
University of Texas at Austin, USA.
Barnden, John A.
1998 An AI system for metaphorical reasoning about mental states in discourse. In
Discourse and Cognition: Bridging the Gap, Jean-Pierre Koenig (ed.), 167–
188. Cambridge: Cambridge University Press.
1999 An implemented system for metaphor-based reasoning, with special applica-
tion to reasoning about agents. In Computation for Metaphors, Analogy, and
Agents, Crystopher Nehaniv (ed.), 143–153. (Lecture Notes in Artificial Intel-
ligence 1562). New York: Springer.
Bateman, John, Bernardo Magnini and Giovanni Fabris
1995 The Generalized Upper Model knowledge base: organization and use. In To-
wards Very Large Knowledge Bases: Knowledge Building and Knowledge
Sharing, Nicolaas Mars (ed.), 60–73. Amsterdam: IOS Press.
Bentley, Peter
1999 Evolutionary Design by Computers. Amsterdam: Elsevier.
Binsted, Kim
1996 Machine humour: An implemented model of puns. PhD dissertation, Univer-
sity of Edimburgh, Scotland.
Boden, Margaret
1990 The Creative Mind: Myths and Mechanisms. London: Weidenfeld and Nichol-
son.
1999 Computational models of creativity. In Handbook of Creativity, Robert Stern-
berg (ed.), 351–373. Cambridge: Cambridge University Press.
Boring, Edwin G.
1923 Intelligence as the tests test it. New Republic 34: 35–37.
van den Bosch, Alexander
2001 Rationality in discovery: a study of logic, cognition, computation and neu-
ropharmacology. PhD dissertation, University of Groningen, Holland.
Bringsjord, Selmer and David A. Ferrucci
2000 Artificial Intelligence and Literary Creativity. Mahwah, New Jersey: Erlbaum.
240 Bibliography

Bulhak, Andrew C.
2000 On the simulation of postmodernism and mental debility using recursive tran-
sition networks. Department of Computer Science Technical Reports, Monash
University, Melbourne, Australia.
Campos, José and A. Dias Figueiredo
2001 Searching the unsearchable: Inducing serendipitous insights. In Proceedings
of the First Workshop on Creative Systems, Amı́lcar Cardoso, Carlos Bento
and Geraint Wiggins (eds.), 159–165. International Conference of Case-Based
Reasoning (ICCBR-01), Vancouver, Canada.
Chandler, Daniel
2002 Semiotics: The basics. London: Routledge.
Cohen, Harold
1981 On the modelling of Creative Behaviour. Rand paper p.6681. Santa Monica,
California: Rand Corporation.
Colton, Simon
2001 Experiments in meta-theory formation. In Proceedings of the AISB’01 Sympo-
sium on AI and Creativity in Arts and Sciences, Geraint Wiggins (Ed.), 100–
109. Brighton: SSAISB.
Colton, Simon, Alan Bundy and Toby Walsh
1999 HR: Automatic concept formation in pure mathematics. In Proceedings of the
International Joint Conference on Artificial Intelligence, Thomas Dean (ed.),
786–793. IJCAI’99. Burlington, Massachussets: Morgan Kaufmann.
2000 Agent based cooperative theory formation in Pure Mathematics. In Proceed-
ings of the Symposium on Creative and Cultural Aspects and Applications of
AI and Cognitive Science, Geraint Wiggins (ed.), 11–18, Birmingham, UK.
Colton, Simon, Alison Pease and Graeme Ritchie
2001 The effect of input knowledge on Creativity. In Proceedings of the First Work-
shop on Creative Systems, Amı́lcar Cardoso, Carlos Bento and Geraint Wig-
gins (eds.), 138–144. International Conference of Case-Based Reasoning (ICCBR-
01), Vancouver, Canada.
Conklin, Darrell and Ian Witten
1995 Multiple viewpoint systems for music prediction. Journal of New Music Re-
search 1(24). London: Routledge.
Cope, David
1991 Computers and Musical Style. Middleton, Wisconsin: A-R Editions.
Costello, Fintan J.
1997 Noun-noun conceptual combination: the polysemy of compound phrases. PhD
dissertation, Trinity College, Dublin, Ireland.
Costello, Fintan J. and Mark T. Keane
2000 Efficient Creativity: Constraint-guided Conceptual Combination. Cognitive
Science 24(2): 299–349. New Jersey: Erlbaum.
Coulson, Seana
2000 Semantic Leaps: Frame-shifting and Conceptual Blending in Meaning Con-
struction. New York/Cambridge: Cambridge University Press.
 Bibliography 241

Cox, Catherine
1926 Genetic Studies of Genius: The Early Mental Traits of Three Hundred Ge-
niuses. Vol. 2. Stanford: Stanford University Press.
Craft, Alastair and Ian Cross
2003 A n-gram approach to fugal exposition composition. In Proceedings of AISB’03
Symposium on AI and Creativity in Arts and Science, Pablo Gervas and Simon
Colton (eds.), 74–83, Brighton: SSAISB.
Csikszentmihalyi, Mihalyi
1996 Creativity: Flow and Psychology of Discovery and Invention. New York: Harper-
Collins.
Eliasmith, Chris and Paul Thagard
2001 Integrating structure and meaning: a distributed model of analogical mapping.
Cognitive Science 25: 245–286. New Jersey: Erlbaum.
Ernst, George and Allen Newell
1969 GPS: A case study in Generality and Problem Solving. New York: Academic
Press.
Evans, Thomas G.
1968 A heuristic program to solve geometric analogy problems. PhD dissertation.
Cambridge, Massachussets: MIT Press.
Falkenhainer, Brian and Kenneth D. Forbus and Dedre Gentner
1989 The Structure Mapping Engine: Algorithm and examples. Artificial Intelli-
gence 41: 1–63. Burlington, Massachussets / Amsterdam: Morgan Kaufmann/Elsevier.
Fauconnier, Gilles and Mark Turner
1998 Conceptual Integration Networks. Cognitive Science 22(2): 133–187. New
Jersey: Erlbaum.
2002 The Way We Think. New York: Basic Books.
Ferguson, Ronald, Kenneth Forbus and Dedre Gentner
1997 On the proper treatment of noun-noun metaphor: A critique of the Sapper
model. In Proceedings of the Nineteenth Annual Conference of the Cognitive
Science Society, Pat Langley and Michael G. Shafto (eds.), 913–914. Hills-
dale, New Jersey: Erlbaum.
Figueiredo, António D. and José Campos
2001 The serendipity equations. In Proceedings of the First Workshop on Creative
Systems, International Conference of Case-Based Reasoning, Amı́lcar Car-
doso, Carlos Bento and Geraint Wiggins (eds.), 121–125. ICCBR-01, Van-
couver, Canada.
Finke, Ronald and Thomas Ward and Steven Smith
1992 Creative Cognition: Theory, Research and Applications. Cambridge: MIT press.
Forbus, Kenneth D. and Daniel Oblinger
1990 Making SME greedy and pragmatic. In Proceedings of the Twelfth Annual
Meeting of the Cognitive Science Society, 61–68. Hillsdale, New Jersey: Earl-
baum.
French, Robert M.
2002 The computational modeling of analogy-making. Trends in Cognitive Sciences
6(5): 200–205. Amsterdam: Elsevier.
242 Bibliography

Gagné, Christina and Edward Shoben

1997 Influence of thematic relations on the comprehension of modifier-noun com-
binations. Journal of Experimental Psychology: Learning, Memory and Cog-
nition 23, 71–87. Washington: APA Journals.
Gentner, Dedre
1983 Structure-mapping: A theoretical framework for analogy. Cognitive Science
7(2): 155–170. Hillsdale, New Jersey: Erlbaum.
Gentner, Dedre and Keneth Forbus
1995 MAC/FAC: a model of similarity-based retrieval. Cognitive Science 19(2):
141–205. Hillsdale, New Jersey: Erlbaum.
Gervás, Pablo
2000 WASP: Evaluation of different strategies for the automatic generation of span-
ish verse. In Proceedings of the AISB’00 Symposium on Creative and Cultural
Aspects and Applications of AI and Cognitive Science, Geraint Wiggins (ed.),
46–65. Brighton: SSAISB.
2001 An expert system for the composition of formal Spanish poetry. Journal of
Knowledge-Based Systems. 14(3-4): 181–188. Amsterdam: Elsevier.
2002 Exploring qualitative evaluations of the creativity of automatic poets. In Pro-
ceedings to the Second Workshop on Creative Systems, Carlos Bento, Amı́lcar
Cardoso and Geraint Wiggins (eds.), 39–47. European Conference of Artifi-
cial Intelligence, ECAI’02, Lyon, France.
Goguen, Joseph
1999 An introduction to Algebraic Semiotics, with applications to User Interface
design. In Computation for Metaphor, Analogy and Agents, Chrystopher Ne-
haniv (ed.), 242–291. (Lecture Notes in Artificial Intelligence 1562). New
York: Springer.
Goldberg, David E.
1989 Genetic Algorithms in Search, Optimization and Machine Learning. Boston,
Massachussets: Addison-Wesley.
Gomes, P., F. C. Pereira, P. Paiva, N. Seco, P. Carreiro, J. L. Ferreira and C. Bento
2002 Case retrieval of software designs using WordNet. In Proceedings of the Eu-
ropean Conference of Artificial Intelligence (ECAI’02), Frank Van Hermelen
(ed.), 245–250. Amsterdam: IOSPress.
Grady, Joe, Todd Oakley and Seana Coulson
1999 Blending and Metaphor. In Metaphor in Cognitive Linguistics, Gerard Steen
and Raymond Gibbs (eds.), 101–125 (Current Issues in Linguistic Theory
175). Amsterdam/Philadephia: John Benjamins
Guilford, Joy P.
1950 Creativity. American Psychologist 5: 444–454. Washington: APA Journals.
1967 The Nature of Human Intelligence. New York: McGraw-Hill.
Hampton, James A.
1987 Inheritance of attributes in natural concept conjunctions. Memory and Cogni-
tion 15: 55–71. Austin, Texas: Psychonomic Society Publications.
1997 Conceptual combination. In Knowledge, Concepts and Categories, Lamberts
and Shanks (eds.), 133–161. London: Psychology Press.
 Bibliography 243

Helmholtz, Hermann Von

1896 Vortäge und Reden. Brunswick, Germany: Friedrich Vieweg.
Hofstadter, Douglas
1995 Fluid Concepts and Creative Analogies. New York: Basic Books.
Hofstadter, Douglas and Melanie Mitchell
1988 Conceptual slippage and mapping: A report of the Copycat project. In Pro-
ceedings of the Tenth Annual Conference of the Cognitive Science Society,
Vimla Patel and Guy Groen (eds.), 601–607. Hillsdale, New Jersey: Erlbaum.
Holyoak, Keith and Paul Thagard
1989 Analogical mapping by constraint satisfaction. Cognitive Science 13: 295–
355. Hillsdale, New Jersey: Erlbaum.
1997 The analogical mind. American Psychologist 52(1): 35–44. Washington: APA
Journals.
Hummel, John and Keith Holyoak
1997 Distributed representations of structure: A theory of analogical access and
mapping. Psychological Review 104: 427–466. Washington: APA Journals.
Johnson, Mark
1987 The Body in the Mind: The Bodily Basis of Meaning, Imagination and Reason.
Chicago: University of Chicago Press.
Karmiloff-Smith, Annette
1993 Constraints on representational change: evidence from children’s drawing.
Cognition 34(1): 47–83. Burlington, Massachussets / Amsterdam: Morgan
Kaufmann/Elsevier.
Keane, Mark T. and Fintan J. Costello
2001 Setting limits on analogy: Why conceptual combination is not structural align-
ment. In The Analogical Mind: A Cognitive Science Perspective, Dedre Gen-
tner and Keith Holyoak and Boicho Kokinov (eds.), 172–198. Cambridge,
Massachussets: MIT Press.
Koestler, Arthur
1964 The Act of Creation. New York: Macmillan.
Kris, Ernst
1952 Psychoanalytic Explorations in Art. New York: International Universities Press.
Lakatos, Imre
1978 Falsification and the methodology of scientific research programs. In Criticism
and The Growth of Knowledge, Imre Lakatos and Alan Musgrave (eds.), 91–
195. Cambridge: Cambridge University Press.
Lakoff, George
1987 Women, Fire, and Dangerous Things: What Categories Reveal About the Mind.
Chicago: University of Chicago Press.
Lakoff, George and Mark Johnson
1980 Metaphors We Live By. Chicago: University of Chicago Press.
Lawson, Anton, Michael Abraham and John Renner
1989 A theory of instruction: Using the learning cycle to teach science concepts and
thinking skills (Monograph). Vancouver: National Association for Research in
Science Teaching (NARST).
244 Bibliography

Lee, Mark and John Barnden

2001 Blending, Analogy and Counterfactuals. In Proceedings of the SEMPRO-01
workshop: Cognitively plausible models of semantic processing. Massimo Poe-
sio (ed.), 67–72 Edinburgh, UK.
Leite, João Alexandre
2003 Evolving Knowledge Bases: Specification and Semantics. Amsterdam: IOS
Press.
Lenat, Douglas
1984 Why AM and EURISKO appear to work. Artificial Intelligence Journal 23:
269–294. Amsterdam: Elsevier.
Levy, Robert
2001 A computational model of poetic creativity with neural network as measure
of adaptive fitness. In Proceedings of the First Workshop on Creative Sys-
tems, International Conference of Case-Based Reasoning, Amı́lcar Cardoso
and Carlos Bento and Geraint Wiggins (eds.), 165–170. ICCBR-01, Vancou-
ver, Canada.
Liu, Hugo and Push Singh
2002 MAKEBELIEVE: Using commonsense knowledge to generate stories. In Pro-
ceedings of the Eighteenth National Conference on Artificial Intelligence (AAAI,
2002), Dechter, Kerns and Sutton (eds.), 957-958. Menlo Park, California:
AAAI Press.
Lothe, Mathis
2000 Knowledge-based composition of minuets by a computer. In Proceedings of
the AISB’00 Symposium on Creative and Cultural Aspects and Applications of
AI and Cognitive Science, Geraint Wiggins (ed.), 25–31. Brighton: SSAISB.
Lubart, Todd I.
1999 Creativity across cultures. In Handbook of Creativity, Robert Sternberg (ed.),
339–351. Cambridge: Cambridge University Press.
Macedo, Luis and Amı́lcar Cardoso
2001 Modelling forms of Surprise in an Artificial Agent. In Proceedings of the
23rd Annual Conference of the Cognitive Science Society, Moore and Sten-
ning (eds.), 588–593. Mahwah, New Jersey: Erlbaum.
Machado, Fernando and Amı́lcar Cardoso
2002 All the truth about NevAr. In Applied Intelligence, Special issue on Creative
Systems, Peter Bentley and David Corne (eds.) 16(2). Dordrecht: Kluwer Aca-
demic Publishers.
Mandelblit, Nili
1997 Grammatical Blending: Creative and schematic aspects in sentence processing
and translation. PhD thesis, UC San Diego
Manurung, Hisar Maruli, Graeme Ritchie and Henry Thompson
2000 Towards a computational model of poetry generation. Informatics Research
Report EDI-INF-RR-0015. Edinburgh, UK.
Marin, Ricardo and Saturnino De La Torre
1991 Manual de la Creatividad. Barcelona: Vicens Vives.
 Bibliography 245

Marshall, James B.
2002 Metacat: A program that judges Creative Analogies in a Microworld. In Pro-
ceedings to the Second Workshop on Creative Systems, Carlos Bento, Amı́lcar
Cardoso and Geraint Wiggins (eds.), 77–84. European Conference of Artifi-
cial Intelligence, ECAI’02, Lyon, France.
Martin, James H.
1990 A computational Model of Metaphor Interpretation. Amsterdam: Academic
Press.
Martindale, Colin
1999 Biological bases of Creativity. In Handbook of Creativity, Robert Sternberg
(ed.), 137–153. Cambridge: Cambridge University Press.
Martindale, Colin and James Greenough
1974 The differential effect of increased arousal on creative and intellectual per-
formance. Journal of Genetic Psychology 124: 311–320. Washington: Heldref
Publications.
Mendes, Mateus
2004 Relacões lexicais na geração de lı́ngua natural. MSc Dissertation. Departa-
mento de Eng. Informática da Universidade de Coimbra. Coimbra, Portugal.
Milosavljevic, Maria and Robert Dale
1996 Strategies for comparison in encyclopedia descriptions. In Proceedings of the
8th. International Workshop on Natural Language Generation, Donia Scott
(ed.), 161–170. INLG’96, Herstmonceux, England.
Mitchell, Tom
1978 Version Spaces: An approach to Concept Learning. PhD thesis. Stanford Uni-
versity.
Murphy, Gregory L. and Douglas L. Medin
1985 The role of theories in conceptual coherence. Psychological Review 92. Wash-
ington: APA Journals.
Neisser, Ulric
1976 Cognition and Reality. San Francisco: Freeman.
Oxman, Rivka
1997 Design by re-representation: a model of visual reasoning in design. Design
Studies 18(4). Amsterdam: Elsevier.
Pazos, A., A. Santos, B. Arcay, J. Dorado, J. Romero and J. Rodriguez.
2002 Application framework to build computer evolutionary systems in Music. LEONARDO
35(4). Cambridge, Massachussets: MIT Press.
Pease, Alison, Daniel Winterstein and Simon Colton
2001 Evaluating Machine Creativity. In Proceedings of the First Workshop on Cre-
ative Systems, International Conference of Case-Based Reasoning, Amı́lcar
Cardoso and Carlos Bento and Geraint Wiggins (eds.), 45–53. ICCBR-01,
Vancouver, Canada.
Peirce, Charles Saunders
1958 Collected Writings from 1931 to 1958. Cambridge, Massachussets: Harvard
University Press.
Pereira, Francisco C.
1998 Modelling Divergent Production: a multi domain approach. In Proceedings
246 Bibliography

of European Conference of Artificial Intelligence, ECAI98, Henri Prade (ed.),

131-132. Hoboken, New Jersey: John Wiley and Sons.
2000 Constru cão interactiva de Mapas Conceptuais (in Portuguese). MSc disserta-
tion, Universidade de Coimbra, Portugal.
2005 Um modelo computacional de Criatividade. PhD Dissertation, Universidade
de Coimbra, Portugal.
Pereira, Francisco C. and Amı́lcar Cardoso
1999 Wondering in a multi-domain environment with Dr. Divago. In Proceedings
of the 8th Cognitive Science and Natural Language Processing Conference
(CSNLP-8), McKevitt, Paul, Conn Mulvihill and Sean O’Nuallain (eds.), 37–
45. Galway, Ireland.
2000 A module for automatic learning of Concept Maps. In Proceedings of Di-
agrams 2000, Anderson, Cheng and Haarslev (eds.), 468–470. New York:
Springer.
2002 The Boat-House visual blending experience. In Proceedings to the Second
Workshop on Creative Systems, Carlos Bento, Amı́lcar Cardoso and Geraint
Wiggins (eds.), 19–25. European Conference of Artificial Intelligence, ECAI’02,
Lyon, France.
2003a The horse-bird creature generation experiment. AISB Journal 1(3): 33–45.
Brighton: SSAISB.
2003b Optimality Principles for Conceptual Blending: A first computational approach.
In Proceedings of AISB’03 Symposium on AI and Creativity in Arts and Sci-
ence, Pablo Gervás and Simon Colton (eds.), 64–73. Brighton: SSAISB
2003c Optimality Principles for Conceptual Blending: A first computational approach.
AISB Journal 1(4). Brighton: SSAISB.
2006 Experiments with free concept generation in Divago (revised and updated).
Knowledge Based Systems 19(7): 459–471. Amsterdam: Elsevier.
Pereira, Francisco C. and Pablo Gervás
2003 Natural language generation from concept blends. In Proceedings of AISB’03
Symposium on AI and Creativity in Arts and Science, Pablo Gervás and Simon
Colton (eds.), 27–33. Brighton: SSAISB.
Pereira, Francisco C., Mateus Mendes and Pablo Gervás and Amı́lcar Cardoso
2005 Experiments With assessment of creative systems: An application of Ritchie’s
criteria. In Proceedings of the IJCAI 2005 Computational Creativity Work-
shop, Pablo Gervás and Alison Pease and Tony Veale (eds.), 50–57. University
of Edinburgh, UK.
Pereira, Francisco C., Ana Oliveira and Amı́lcar Cardoso
2000 Extacting concept maps with Clouds. In Lecture Notes in Artificial Intelli-
gence. Proceedings to the Argentine Symposium on Artificial Intelligence (ASAI’00),
Gabriela Henning (ed.), 60–68. New York: Springer.
Plate, Tony A.
1994 Distributed representations and nested compositional structure. Ph.D. disser-
tation. University of Toronto, Canada.
Popper, Karl
1959 The Logic of Scientific Discovery. London: Hutchinson & Co.
 Bibliography 247

Ram, Ashwin, Linda Wills, Eric Domeshek, Nancy Nersessian and Janet Kolodner
1995 Understanding the Creative Mind. Artificial Intelligence Journal, 79: 111–
128. Amsterdam:Elsevier.
Reffat, Rabee M.
2002 Intelligent agents for concept invention in design. In Proceedings of the Inter-
national Workshop on Agents in Design (WAID’02), John Gero and Frances
Brazier (eds.), 55-68. Cambridge, MA.
Ribeiro, P. and F. C. Pereira and M. Ferrand and A. Cardoso
2001 Case-based melody generation with MuzaCazUza. In Proceedings of the AISB’01
Symposium on AI and Creativity in Arts and Sciences, Geraint Wiggins (ed.),
67-74. Brighton: SSAISB.
Ribeiro, P., F. C. Pereira, B. Marques, B. Leitão and A. Cardoso
2003 A model for Creativity in creature generation. In Proceedings of the 4th In-
ternational Conference on Intelligent Games and Simulation (GAME ON’03),
Qas Mehdi and Norman Gough (eds.), 175–180. EuroSIS / University of Wolver-
hampton, UK.
Ritchie, Graeme D.
2001 Assessing Creativity. In Proceedings of the AISB’01 Symposium on AI and
Creativity in Arts and Sciences, Geraint Wiggins (ed.), 3–11. Brighton: SSAISB.
Rohrer, Tim
2000 Even the interface is for sale: Metaphors, visual blends and the hidden ide-
ology of the internet. In Language and Ideology: Cognitive Descriptive Ap-
proaches, René Dirven and Roslyn Frank and Cornelia Ilie (eds.), 189–214.
Amsterdam/Philadelphia: John Benjamins.
Romero Cardalda, Juan Jesus
1999 Metodologı́a evolutiva para la construcción de modelos cognitivos complejos.
Exploración de la ‘creatividad artificial’ en composición musical. PhD Disser-
tation. Universidad de La Coruña, Spain.
Rosch, Eleanor
1975 Cognitive representations of semantic categories. Journal of Experimental Psy-
chology: General 104: 192–232. Washigton: APA Journal.
Saunders, Rob and John Gero
2001 The Digital Clockwork Muse: A computational model of aesthetic evolution.
In Proceedings of the AISB’01 Symposium on AI and Creativity in Arts and
Sciences, Geraint Wiggins(Ed.), 12–21. Brighton: SSAISB.
De Saussure, Ferdinand
1983 Course in General Linguistics. London: Duckworth.
Sims, Karl
1991 Art Evolution for computer graphics. ACM Computer Graphics 25. New York:
ACM Press.
Sosa, Ricardo and John Gero
2003 Design and Change: A model of Situated Creativity. In Proceedings of the
Third Workshop on Creative Systems, Amı́lcar Cardoso, Carlos Bento and
John Gero (eds.) IJCAI-03, Acapulco, Mexico.
248 Bibliography

Steel, Graham
1999 Cross-domain Concept Formation using HR. MSc Thesis. University of Edin-
burgh, UK.
Sternberg, Robert and Todd Lubart
1996 Investing in Creativity. American Psychologist 51: 677–688. Washington: APA
Journals.
1999 The Concept of Creativity: Prospects and Paradigms. In Handbook of Creativ-
ity, Robert Sternberg (ed.), 3–16. Cambridge: Cambridge University Press.
Stock, Oliviero and Carlo Strapparava
2003 Getting serious about the development of Computational Humor. In Proceed-
ings of the Eighteenth International Joint Conference on Artificial Intelli-
gence, Georg Gottlob and Toby Walsh (eds.), 59–64. Burlington, Massachus-
sets: Morgan Kaufmann.
Sweetser, Eve and Barbara Dancygier
2005 Mental Spaces in Grammar: Conditional Constructions. Cambridge: Cam-
bridge University Press.
Thomas, Nigel
1999 Are Theories of Imagery Theories of Imagination? An Active Perception ap-
proach to Conscious mental content. Cognitive Science 23: 207–245. New
Jersey: Erlbaum.
Turner, Mark
2006 Compression and Representation. Language and Literature Journal 15(1):
17–27. London: SAGE Publications.
Veale, Tony
1995 Metaphor, Memory and Meaning: Symbolic and Connectionist issues in Metaphor
interpretation. PhD Dissertation. Dublin City University, Ireland.
Veale, Tony and Mark Keane
1997 The Competence of sub-optimal theories of Structure Mapping on Hard Analo-
gies. In Proceedings of the International Joint Conference on Artificial Intelli-
gence (IJCAI’97), Martha Pollack (ed.), 232–237. Burlington, Massachussets:
Morgan Kaufmann.
1998 Principal differences in Structure Mapping: The contribution of non-classical
models. In Proceedings of Advances in Analogical Research, Keith Holyoak
and Dedre Gentner and Boicho Kokinov (eds.). New University of Bulgaria
Press.
Veale, Tony and Diarmuid O’Donoghue
2000 Computation and Blending. Cognitive Linguistics 11 (3): 253–281 (Special
Issue on Conceptual Blending). Berlin/New York: Mouton de Gruyter.
Wallas, Graham
1926 The Art of Thought. New York: Harcourt-Brace.
Ward, Thomas, Steven Smith and Ronald Finke
1999 Creative Cognition. In Handbook of Creativity, Robert Sternberg (ed.), 189–
213. Cambridge: Cambridge University Press.
Weisberg, Robert
1999 Creativity and Knowledge: A challenge to theories. In Handbook of Creativity,
Robert Sternberg (ed.), 226–251. Cambridge: Cambridge University Press.
 Bibliography 249

Wiggins, Geraint A.
2001 Towards a more precise characterisation of creativity in AI. In Proceedings of
the First Workshop on Creative Systems, International Conference of Case-
Based Reasoning (ICCBR-01), Amı́lcar Cardoso, Carlos Bento and Geraint
Wiggins (eds.), 113–129. Vancouver, Canada.
2003 Categorising creative systems. In Proceedings of the Third Workshop on Cre-
ative Systems, Amı́lcar Cardoso, Carlos Bento and John Gero (eds.), 4–13.
IJCAI-03, Acapulco, Mexico.
Winston, Patrick H.
1980 Learning and reasoning by analogy. Communications of the ACM 23(12):
689–703. New York: ACM Press.
Wisniewski, Edward J.
1997 Conceptual Combination: Possibilities and aesthetics. In Creative thought.:
An investigation of conceptual structures and processes, Thomas Ward and
Steven Smith and Jyotsna Vaid (eds.), 51–81. Washington: American Psycho-
logical Association.
Index

algebraic semiotics, 63, 117 concept, 47

Amabile, 27 discovery, 49
Analogy, 68 exemplar view, 47, 142
ATT-Meta, 79 formation, 47, 49
invention, 47, 50, 99, 101
Bisociation, 14, 85, 92, 97, 101, 206 invetion, 89
bisociative mechanism, 90 learning, 49
blending examples micro-theories, 48
CEO boxing fight, 185, 187, 188, 213 micro-theory, 104, 107, 142
Computer desktop, 185, 187, 219 prototype view, 47
Computer virus, 185, 187 theory view, 48
Gun wound, 185, 187, 215 Conceptual Blending, 14, 53, 55, 92, 115,
Kant debate, 185, 187, 214 174, 206
Land yacht, 185, 187, 215 blend, 56, 110
Mabel is the daughter of Paul, 185, completion, 58
187, 220 completion evidence, 131
Pet fish, 185, 187, 215 composition, 58, 120
Same-sex marriage, 185, 187, 213 conceptual integration, 53
Sandwich counterfactual, 185, 187, counterfactuals, 65
220 cross-space mappings, 56
Trashcan basketball, 185, 187, 213 double-scope blend, 62, 190
boat-house, 142 elaboration, 59
Boden, 18 elements, 56
combinatorial creativity, 24 emergent structure, 59
conceptual space, 24 frame, 57, 58, 127
e-creativity, 24, 25 framework, 55
exploratory-transformational Creativ- mental space, 56
ity, 24 optimality principles, 61, 102, 122,
h-creativity, 23, 30 126, 206
p-creativity, 24, 30 Elaboration, 136
t-creativity, 24, 25, 30 Integration, 61, 127, 128, 155
taxonomy, 23 Intensification of Vital Relations,
Boring, 8 62, 132, 237
Maximization of Vital Relations,
Case-Based Reasoning, 42, 149 62, 132, 155
Catherine Cox, 8 Pattern Completion, 61, 131, 155
central limit theorem, 141 Relevance, 62, 136, 155
Clouds, 105, 142 Topology, 61, 130, 155, 237
cognition-centered model, 42 Unpacking, 62, 131, 133, 155, 237
comparison to an archetype, 150 Web, 62, 135
compression, 133 selective projection, 58
Computational Creativity, 30 single-scope blend, 62, 190
 Index 251

Conceptual Combination, 51, 68, 92, 161 main concept, 104

C3 , 53, 161, 171 relations, 103, 105
diagnosticity, 52, 171 concept micro-theory, 112
informativeness, 52, 171 Constraints, 102, 126
plausibility, 52, 171 domain, 104
conjunctive, 51 Elaboration, 102, 184
hybrid interpretations, 163 element, 104
known-concept, 51 exhaustive closure, 105
property interpretation, 163 Factory, 102, 122
property mapping, 51 frames, 101, 103, 108, 127, 155
relational, 51 organizing, 109
Conceptual Combination theories pattern identifying, 109
Abstract relations, 51 transforming, 109
Composite Prototype, 51, 236 generic domain, 114, 165
Constraints, 51, 52 instances, 101, 103, 113, 231
Dual-Process, 51 integrity constraints, 101, 103, 107,
convergent thinking, 22 111, 179
Copycat, 82, 201 Interpreter, 160, 181
creature generation, 173 Knowledge Base, 101, 103
creatures, 173 Mapper, 101, 115, 163, 184
cross-domain, 96 mapping operation, 118
cross-domain transfer, 11 projections, 101
Csikszentmihalyi, 11, 19 rules, 101, 103
creativity traits, 20 selective projection, 123
creativity with a big C, 19, 98 syntax
domain, 19 frame/6, 108
elaboration, 21 integrity/3, 112
evaluation, 21 Prolog scripts, 237
field, 19 rule/6, 108
incubation, 21 transfer operation, 124
divergent thinking, 21, 22, 27, 28, 86, 206,
insight, 21
235
person, 19
domain-centered model, 42
preparation, 21
Drama, 81
Dupond, 200
defining frame, 134
Divago, 91, 100, 101, 206 edit distance, 150
architecture, 101 Elaboration, 90
blend, 120 embodiment, 14
Blender, 101, 117 established blends, 165
blending projection, 119 evaluation, 90
blendoid, 121 evolutionary model, 42
complexity, 116, 122 experiments, 141
concept, 104 Expert Systems model, 42
concept map, 101, 103, 105, 107, 118
concepts, 103 fine-tuning, 159, 211
elements, 105 focal concept, 164
252 Index

frame coverage, 127 Humour, 15

frames, 127, 223 matrices of thought, 14
free generation, 166 Science, 16
strategies, 14
games, 173 The Act of Creation, 14, 60
Geneplore model, 26, 93
preinventive structures, 26 level of arousal, 27
General Problem Solver, 87 LISA, 80
Generalized Upper Model, 104 Logo, 113, 143, 231
Genetic Algorithm, 102, 123, 149
crossover, 124 MAC/FAC, 208
fitness function, 124, 157, 167 Magic The Gathering, 174, 178
genotype, 124 meta-level reasoning, 40, 85, 100
mutation, 124 Metaphor, 68
phenotype, 124 Model of Concept Invention, 85, 206
population, 124 alphabet, 91
reproduction, 124 convergent strategy, 96, 97, 139
roulette-wheel, 125 cross-concept based elaboration, 96,
goal, 90 137
goal frames, 136 cross-domain based elaboration, 96
Guilford, 8 cross-domain mapping function, 92
cleverness, 13 divergent strategy, 90, 95, 97, 139
Convergent Production (CP), 11 elaboration function, 95
Creative Production, 10 emergence, 92
Divergent Production (DP), 8, 10, 206 empty concept, 92
elaboration, 10 internal logic elaboration, 96, 137
flexibility, 10 Knowledge Base, 91
fluency, 10 language, 91
Creativity Model, 13 mapping function, 92
originality, 10 novelty, 94
Structure Of Intellect (SOI), 8 rule sets, 91
transfer recall, 11 rule-based elaboration, 96, 137
strategy, 94, 97
head noun, 164 transfer operation, 93, 99, 118
Helmholtz, 8 universe, 91, 96
horse-bird, 149 usefulness, 94
house-boat instances, 231 multi-domain environment, 12, 201
HR, 39–41, 49, 212 multi-domain knowledge base, 89

image schema, 111 Natural Language Generation, 160

inner-space relations, 133 normalized minimum distance, 150
Investment theory, 27 noun-noun combinations, 161
novelty, 29, 94, 99, 126, 148, 151, 161
Koestler, 14, 202
Arts, 17 outer-space relations, 133
Bisociation, 14
codes of rules, 14 Pegasus, 151, 154, 157
 Index 253

poetry generation, 200 usefulness, 29, 94, 99, 126, 151, 161

re-representation, 28, 85 Veale

reasoning mechanism, 90 Conceptual Scaffolding, 71
reinvention, 211, 237 Sapper, 41, 65, 71, 116
Ritchie, 32, 158, 170, 181, 199, 207, 237 dormant bridges, 75, 116
basic item, 33 incremental rule, 236
insipiring set, 32 opportunistic activation flow, 75
inspiring set, 211, 237 squaring rule, 75
measures, 34 structural inference, 75
rating schemes, 33 triangulation rule, 75, 116
typicality, 33, 94, 150, 236 visual morphing, 146
value, 33, 236 vital relations, 133

serendipity, 32, 35 Wallas, 8

single frame integration, 127 WASP, 199, 212
SME, 69 Weisberg, 28
space of blends, 148 Wiggins, 31, 91, 94, 207
stochastic model, 42 conceptual space, 31
structure alignment, 56, 83, 117, 130, 163, traversal strategy, 31, 91
176, 201 universe, 32, 91
Systems model, 19, 41

tuning, 164, 172