Serial verb constructions and covert coordinations in Edo

An analysis in a dynamic frame theory

Data and central issue. Edo is a language spoken in Central Nigeria and belongs to the Kwa-family, in which
serial verb constructions (SVCs) are a characteristic part of the grammar. Summarizing the existing consensus,
[Aik06] defines an SVC as a monoclausal sequence of verbs which act together as a single predicate, without
any overt marker of coordination or subordination and a single value for tense and aspect. For Edo, Aikhenvald’s
definition must be strengthened in two respects: SVCs have only one subject and at least one internal argument is
shared by the verbs. A central difference in Edo is that between a consequential SVC and a covert coordination
(CC), [Ste01]. Consider the following two examples taken from [BS99, p.3].
(1) a. Òzó ghá gbè è.wé khié.n.
Ozo FUT hit goat sell
‘Ozo will kill the goat and sell it.’ Consequential SVC (CSVC)
b. Òzó ghá gbè è.wé khié.n ùhùnmwùn éré.n.
Ozo FUT hit goat sell head its
‘Ozo will kill the goat and sell its head.’ Covert Coordination (CC)
In a CSVC the verbs are either transitive or ditransitive. The subjects and direct objects are always identified with
each other, i.e. they are coreferential. By contrast, an indirect object of a ditransitive verb is never identified with
any argument of the other verb. In particular, the indirect objects are not identified if both verbs are ditransitive.
(2) Úyi hàé Ìsò.kè.n ı́ghó dó-rhié.
Uyi pay Isoken money steal
‘Uyi paid Isoken the money and stole it.’ [Ste01, p.137]
Despite the fact that the subjects are always identified, it is not possible to have a subject pronoun before V2 in a
CSVC, (3-a). This restriction does not hold for a CC, as shown by (3-b).
(3) a. Òzói mú èmá (*Ó.i ) kpèé..
Ozo carry drum he beat [Ste01, p.64]
b. Òzói gbó.ó. ı́vı̀n Ó.i bòló ó.kà.
Ozo plant coconut he peel corn
‘Ozo planted coconut and he peeled the corn.’ [Ste01, p.65]
If the object arguments in a CC are coreferential, there is a pronoun after the second verb, (4-a). For a CSVC, such
a pronoun is not admissible. (4-b) can only be interpreted as a CC, having the same meaning as (4-a).
(4) a. Òzói lé ı́zè.j Ó.i rrı́ ó.rèj .
Ozo cook rice he eat it
‘Ozo cooked rice and he ate it.’ [Ste01, p.65]
b. *Òzó le izei rri orei .
Ozo cook rice eat it [Ste01, p.65]
Semantically, CSVCs and CCs differ in the following respect. For a CSVC, the action expressed by the first verb
is done with the intention to carry out an action expressed by the second verb. For example, (1-a) can only be true
if Ozu killed the goat with the intention of selling it afterwards. If he killed the goat by accident or decided to sell
it only after the killing, (1-a) is false. No corresponding restriction exists for a CC. For example, (4-a) is true if
Ozu cooked the rice and ate it afterwards.
Informal outline of the analysis: basic assumptions. The semantic theory to be presented in this talk is based
on the following assumptions: (i) the semantic relation expressed by CSVCs and CCs in Edo is based on coherence
relations which, in turn, are defined as complex relations between events; (ii) these relations are defined in terms
of typed attribute-value pairs in a frame theory which is used to semantically model proper names, common
nouns and verbs. CSVCs CCs and coherence relations. The semantic characterization of CSVCs and CCs in the
previous section has shown that they differ in the way the events described by the verbs in both constructions are
related to each other. These differences are located at at least two different levels: mereological and (constraints
on) participants. Underlying the semantic interpretation of a CSVC is a plan (see [Bit01]). The events described
by the verbs in this construction are part of an intended plan, given by an event e, which consists
F of n component
events e1 , . . . , en as its constituent (material) parts s.t. ei v e and ei < ej for i < j and ei = e. Hence, an
event e is related to two are more events which are linearly ordered and each of which is a material part of e.
By contrast, for a CC, only two events are related s.t. the first temporarily precedes the second and, therefore,
the two events are mereologically disjoint. In addition to these differences at event structure there is a difference
w.r.t. how the participants in the events are related to each other. Whereas in the case of a planned event the actor
and the theme are always identical, for a CC only the actors are required to be coreferential. This characterization
suggests modelling the relation expressed by CSVCs and CCs in terms of coherence relations (CR). One way of
modelling such relations is in terms of complex relations between events. We use frame theory to implement this
idea. In particular, we assume that event frames for verbs in Edo have an attribute COHERENCE RELATION which
specifies a possible relation to the next event described in the text (discourse). The values of this attribute are
of type plan and list. Each type has the attribute MEREOLOGICAL RELATION which specifies the mereological
relation between the event at the root of the frame and the current topic event. For a CR of type plan, the former
is a proper material part of the latter whereas in the case of a CR of type list the two are disjoint. Constraints on
participants are modeled by requiring the corresponding thematic relations to have the same values.
Coherence relations trigger expectations. After processing the first VP in a CSVC or a CC, a comprehender
does not yet know what the relation to the next event will be. The event described by this part can either be a
part of a larger plan as in (1-a) where the killing is done with the intention of selling the goat, or it simply is
the first event in a succession of events with a shared actor, as in (1-b). However, (s)he knows that this relation
will be exactly one of these two possibilities. This knowledge is used by her/him to non-deterministically extend
the current information state with those three possibilities so that the next part of the construction is interpreted
in relation to this event. How can a decision be made between the two possibilities? Pronouns as indicators of
coherence relations. Recall from the first section that in a CSVC shared arguments are never overtly realized by a
pronoun on the n-th verb for n > 1. Indirect objects of ditransitive verbs are realized by proper names or common
nouns since they can never be coreferential. By contrast, in a CC coreference of the actor role can be marked
by a pronoun and coreference of theme roles is marked by a pronoun after the second verb. Hence, a pronoun is
an indicator of a list scenario and excludes a plan scenario. This is similar to the way a word in English can be
an indicator of a particular coherence relation in English. For example, [KR13] found that in a context ‘Amanda
amazed Brittany because . . . ’ with the implicit causality verb ‘amaze’ the connective ‘because’ is an indicator
of an explanation relation since it raises the probability for this coherence relation to (almost) 1. Hence, if after
processing say ‘Ozu le ize’ a pronoun is encountered (‘O’), the comprehender knows that the speaker describes a
list scenario and not a scenario of type plan. If no pronoun is encountered, both types of coherence relations (plan
as well as list) remain options. It is only the region after the second verb which allows for a decision. If a pronoun
is encountered, the scenario described must be of type list, and of type plan otherwise. Thus, one function of a
pronoun consists in eliminating possibilities related to coherence relations.
Outline of the formalization. We define a probabilistic dynamic update semantics with frames. Models. A
probabilistic world model with frames is a tuple hW, D, I, {fd,σ,w }d∈D,σ∈Σ,w∈W i. W is a finite set of possible
worlds which is used to represent (epistemic) uncertainty. An example is the uncertainty w.r.t. to the value of the
COHERENCE RELATION -attribute in event frames. The domain D = {Dσ }σ∈Σ is the union of finite domains
Dσ based on a partially ordered sort hierarchy hΣ, vΣ i with basic sorts like ‘event’ (e) or ‘individual’ (d). D is
structured by a (material) part relation v. F = {fd,σ,w }d∈D,σ∈Σ,w∈W is the domain of frames. Each frame is of a
sort σ and is a (generated) submodel of a possible world w, namely, the information associated with a particular
object d in that world which is the root of the frame. Fw is the set of frames in world w and fw,d is the frame
in w with root d. A frame is related to a set of relations on D × D. Each relation R corresponds to a finite path
(chain of attributes) ≥ 0 starting at the root d. The domain of R is given by the source-sort of the first attribute
in the path and the range of R by the target-sort of the last attribute in the path. For path of length 0, one defines
the shift: λQσ .λx.λy.Qσ (x) ∧ x = y. Each R must always be satisfied at the root. Hence, a frame fw with root
d corresponds to a complex property Qfw = Q0 ∩ Q1 ∩ . . . · Qn s.t. each Qi is the domain of a relation R and
one has Qfw (d) is true in w iff Qi (d) is true in w for each 1 ≤ i ≤ n. Using this fact, we define a relation θ s.t.
Q ∈ θ(f )(d) iff Q(w)(d). θ(f )(d) = σ means that Qf (d) is of type σ. θ(f )(d)π denotes the set of properties
which hold at the end of path π in the frame f with root d.
Information states in a frame theory. An information state s consists of a set of possibilities i. A possibility
i consists of a world w, two stacks (following Incremental dynamics, [vE01]) and two functions γ1 and γ2 . The
stack c1 assigns values to discourse parameters which are variable. Examples are speech time, speaker, hearer
etc. In the present context we are interested in the parameter ‘topic event’, which is assumed to be located at the
0-th position of c1 . The stack c2 consists of those objects which are introduced by common nouns, proper names
and verbs. The function γ assigns to each element o of c2 a frame f ∈ Fw s.t. o is the root of f . We define two
projection functions (see also [Bit01]): pi which yields the i-th element on a stack counted from the top of the
stack, i.e. p0 (c) is the top element of c. The projection function pσ yields the restriction of the stack to objects of
type σ. pi (pσ (c)) is the i-th object of type σ on stack c. The distinction between the topic event, which belongs
to c1 , and the current top-most event, which is an element of c2 , is motivated by the following reasons: (i) in case
of a plan scenario the topic event is the planned event e which remains constant while its component events are
(successively) introduced on the stack c2 ; by contrast, in a list scenario the topic event is changed with each new
verb because the events are not related at the mereological level; (ii) it is used to implement the mereological
relation between two events in a plan and list scenario and (iii) arguments provide information about the top-most
event on the c2 stack, independently of the mereological relation to other events either. The functions γ1 and γ2
assign to each element on c1 and c2 its frame fw in the world w of the possibility i.
Update operations. We provide simplified versions of the most important update operations The difference
between the way commons (and proper names) and verbs function is reflected in having two update operations.
Update operations for common nouns and proper names. The update operation for cn’s and pn’s s[d] is a do-
main extension operator, similar to s[x] in other update semantics. The difference lies in the fact that each element
on the stack is paired with a frame. The definition is s[d] = {hw0 , c01 , c02 , γ10 , γ20 i | ∃n∃i = hw, c1 , c2 , γ1 , γ2 i ∈
s ∧ w = w0 ∧ γ10 = γ1 ∧ c01 = c1 ∧ c02 = c2ˆd for d ∈ Dd ∧ γ20 (c02 [i]) = γ2 (c2 [i]) for 0 ≤ i ≤ n − 1 ∧ n =
|c2 | ∧ γ20 (c02 [n]) = fw,d }.
Update operation for events in CSVS and CCs. The combination of two verbs or clauses in an SVC or a CC is
modeled as an update operation (compare the interpretation of ‘.’ in dynamic semantics as function composition of
information states: λpλqλsλs0 .∃s00 .p(s)(s00 ) ∧ q(s00 )(s0 )). The update operation is a conditional one: it extends the
stack c2 by an event which is required to satisfy the value of the COHERENCE RELATION-attribute of the (so far)
top-most event at the root of its frame and it changes the topic event depending on the type of the CR. Constraints
between stack elements are expressed using θ. For example, θ(γ2 (c2 [n]))(c2 [n]) = θ(γ2 (c2 [n − 1]))(c2 [n −
1])(COHERENCE RELATION) requires the newly introduced event to have the same value as the value of the CO -
HERENCE ATTRIBUTE of the event at the previous position. θ(γ1 (c1 [0]))(c1 [0])( MEREOLOGICAL RELATION ) =
θ(γ2 (c2 [n]))(c2 [n))(COHERENCE RELATION)(MEREOLOGICAL RELATION) requires the relation between the topic
event and the top-most event to respect the mereological relation set up by the COHERENCE RELATION-attribute.
The update operation is defined as follows: s[e] = {hw0 , c01 , c02 , γ10 , γ20 i | ∃n∃m∃i = hw, c1 , c2 , γ1 , γ2 i ∈ s ∧ w =
w0 ∧ c02 = c2ˆe for d ∈ De ∧ γ20 (c02 [i]) = γ2 (c2 [i]) for 0 ≤ i ≤ n − 1 ∧ n = |c2 | ∧ m = |c1 | ∧ γ20 (c02 [n]) = fw,e ∧
θ(γ2 (c2 [n]))(c2 [n]) = θ(γ2 (c2 [n−1]))(c2 [n−1])(COHERENCE RELATION)∧θ(γ1 (c1 [0]))(c1 [0])(MEREOLOGICAL RELATION

θ(γ2 (c2 [n]))(c2 [n))(COHERENCE RELATION)(MEREOLOGICAL RELATION) ∧ if θ(γ2 (c2 [n]))(c2 [n]) = list ∧
θ(γ2 (c2 [n − 1]))(c2 [n − 1])(COHERENCE RELATION) = list then c01 [0] = c02 [n] ∧ c01 ([i]) = c1 ([i]) ∧ γ10 (c01 [0]) =
γ20 (c2 [n]) ∧ γ10 (c01 [i]) = γ1 (c1 [i]) f or i 6= 0 else c01 = c1 ∧ γ10 = γ1 }.
Update operation for pronouns. The update operation for pronouns is the sequential composition of three
update operations. Similar to DRT, pronouns push a new object on c2 . this is modeled by s[d]. Next, it is tested
whether the relation between the top-most and the previous event is of type list: s[CR] = {hw0 , c01 , c02 , γ10 , γ20 i | ∃m∃i =
hw, c1 , c2 , γ1 , γ2 i ∈ s : w = w0 ∧ c01 = c1 ∧ m = |c1 | ∧ γ10 = γ1 ∧ c02 = c2 ∧ γ20 = γ2 ∧ θ(γ2 (c2 [n −
1]))(c2 [n − 1]) = list ∧ θ(γ2 (c2 [n − 2]))(c2 [n − 2])(COHERENCE RELATION) = list}. In the probabilistic
setting
P of [KR13] s[CR] implements the raising of pr(CR) to 1 in the equation pr(pronoun = referent) =
CR∈CRs pr(CR) · pr(pronoun = referent|CR). This update operation, therefore, has the effect of eliminat-
ing one particular type of value from the COHERENCE RELATION attribute. The third update operation uses the
fact that a pronoun, being an argument, is always related to a particular thematic relation TR. The constraint im-
posed by s[T R] requires the values of the TR-attribute of the current event and the previous event to be the same:
s[T r] = {hw0 , c01 , c02 , γ10 , γ20 i | ∃i = hw, c1 , c2 , γ1 , γ2 i ∈ s : w = w0 ∧ c01 = c1 ∧ c02 = c2 ∧ γ10 = γ1 ∧ γ20 =
γ2 ∧ θ(γ2 (c2 [n − 1]))(cs [n − 1])(TR) = θ(γ2 (c2 [n − 2]))(cs [n − 2])(TR)}. This implements the fact that in Edo
using a probabilistic setting like that of [KR13] pr(pronoun = referent|list) is 1 for a pronoun in a given argument
position.

[Aik06] Alexandra Aikhenvald. Serial verb constructions in typological perspective. In Alexandra Aikhenvald and R.M. Dixon, editors,
Serial verb constructions, pages 1–68. Oxford University Press, 2006.
[Bit01] Maria Bittner. Surface composition as bridging. Journal of Semantics, 18(2):127–177, 2001.
[BS99] Mark Baker and Osamuyimen Thompson Stewart. On double headedness and the anatomy of the clause. ms. Rutgers University,
1999.
[KR13] Andrew Kehler and Hannah Rohde. A probabilistic reconciliation of coherence-driven and centering-driven theories of pronoun
interpretation. Theoretical Linguistics, 39(1-2):1–37, 2013.
[Ste01] Osamuyimen Thompson Stewart. The serial verb construction parameter. Garland, 2001.
[vE01] Jan van Eijck. Incremental dynamics. Journal of Logic, Language and Information, 10(3):319–351, 2001.