An Account of Event Composition for Indo

An Account of Event Composition for Indo-Aryan Compound Verbs
by
Soma Paul
in
5th International Conference on Generative Approaches to the Lexicon
Report No: IIIT/TR/2009/224
Centre for Language Technologies Research Centre
International Institute of Information Technology
Hyderabad - 500 032, INDIA
September 2009
An Account of Event Composition for Indo-Aryan
Compound Verbs
Soma Paul
Language Technology Research Center
International Institute of Information Technology
Hyderabad, India
[email protected]
Abstract
The paper proposes a complex event
composition mechanism, a variant of cocompositionality (Pustejovsky (1995)), for
building the semantic structure of compound
verbs in Indo-Aryan languages. The
motivation for giving a co-compositional
account for Indo-Aryan compound verbs is
drawn from Langacker’s concept of baseprofile distinction (Langacker 1987) stated
in his cognitive model of perception. The
paper examines how the notion of profiling
which seems to explain very well the
semantic composition of compound verbs
can be represented in a computationally
tractable model.
1.0 Introduction
The paper presents a complex event
composition mechanism that builds the
semantic structure of verbal compounds in
Indo-Aryan languages. The mechanism is a
variant of co-compositionality as proposed
by Pustejovsky (1995). Compound verbs
are composed of a main verb and a light
verb. The light verb has so far mostly been
treated as an auxiliary and the argument
structure list of auxiliary includes a
predicative word (in this case the main verb)
and its arguments (Hinrichs and Nakazawa
(1990), Butt(1995), Alsina (1996)). This
paper argues that light verbs are not
semantically empty and that both main verb
and light verb play a co-operative role in
building the semantics of compound verbs.
The event composition is constrained at the
level of semantics.
Verbs denote situation which can be taken
as a base or frame. A base has a close
resemblance to Goldberg (1995) idea of
construction. More than one verb might
share a situation as a base/frame. However,
depending on which part of the situation is
profiled, the meaning of an individual verb
is determined. Language might use various
syntactic or morphological devices for
achieving the profiling. Light verb in IndoAryan language is one such device. Light
verbs that unify with a main verb highlight a
subpart of that base which can be called
‘profile’. The profile achieves a special
degree of prominence and becomes the
meaning of the compound verb. The paper
examines how can the notion of profiling
1
(which explains very well the semantic
composition of compound verbs) be
represented in a computationally tractable
model. The motivation for such an account
is drawn from Langacker’s concept of baseprofile distinction (Langacker 1987) stated
in his cognitive model of perception. In the
context of Indo-Aryan compound verbs the
issues are manifold:
i.
Although compound verbs mainly
preserve the core meaning of the main
verb, profiling by light verbs can result in
modification of the semantics of
compound verbs so much so that the
syntactic behavior of the resultant
compound becomes different from that of
its main verb counterpart. Light verbs are
therefore not merely auxiliaries. A
semantic composition principle is required
that allows both verbs co-compositely
specify the semantics and syntactic
behavior of compound verbs.
ii.
Light verbs are semantically related
to their independent counterpart. When a
verb occurs as a light verb, it undergoes
some amount of semantic bleaching. The
challenge is to have the light verb derived
from its full verb counterpart with having
a focused semantic feature specification
that is relevant for the light verb variant
and shadowing the rest of information
associated to the full verb counterpart.
iii.
Every Indo-Aryan language has a
fixed set of light verbs. A main verb does
not select all light verbs that are there in a
language. The details of selectional
restrictions have been worked out in Paul
(2004). An account of the compound verb
should necessarily be able to define these
selectional restrictions.
The next section describes the characteristic
features of Indo-Aryan compound verbs.
Section 3 postulates a few semantic types of
verbs and represents them in a multi-layered
representation. Section 4 defines the rules
of co-composition for constituting the
lexical structure of legitimate compound
verbs.
2.0 Data
Compound verbs (henceforth CV) in IndoAryan languages are composed of two verbs,
the first member (V1) is either a participial
form as in Bangla, Odiya and Assamese 1
(see 1a) or a root (as in Hindi2, see in (1b)
and the second member (V2), which is
“semantically bleached”, bears the inflection.
For example,
Bangla
1. meeta heš-e uth-lo
girl-cl laugh-pf rise – 3 pt
‘The girl burst into laughter’
Hindi
2. ləDkI həs pəD-I
girl laugh fall – 3 pt
‘The girl burst into laughter’
Every Indo-Aryan language has a fixed set
of V2s. In this paper we discuss two light
verbs deoa ‘give’ and neoa ‘take’ which are
found in most Indo-Aryan languages. Here
the data is cited from Bangla.
Compound verbs preserve the core meaning
of the main verb. For example, let us
consider Bangla transitive verbal root
banano 3 ‘build’. This verb combines with
the V2s: deoa ‘give’, neoa ‘take’, phæla
1
This paper presents a co-composition
account
(Pustejovsky
1995)
that
dynamically builds the semantics of
compound verbs, which is basically a
function derived from light verbs profiling
the semantics of main verbs. Verbs will be
classified in a hierarchy of semantic types in
order to prevent any illicit derivation of
compound verb.
Bangla, Odiya and Assamese are languages spoken
in eastern and north eastern zone of India. Bangla is
also the official language in Bangladesh. A detailed
study of compound verbs in these languages can be
obtained from Dasgupta (1977), Mohanty(1992),
Paul(2004).
2
Hindi is a widely spoken language in Northern and
central India. The compound verbs of Hindi have
been studied by Hook (1974), Abbi(1991), Butt(1995)
to name a few.
3
The verbal noun form of verbs is used here as the
citation form.
2
‘drop’ and rakha ‘keep’ as shown below in
column 3. The semantic overtone that the
V1
‘build’
banano
V2
CV
Semantic overtone
deoa ‘give’
banie deoa
‘build (a thing for
someone)’
Effect of the action directed towards a
participant other than the actor
banano
neoa ‘take’
banano
Rakha ‘keep’
banano
V2s add to the meaning of banano ‘build’ is
summarized in column 4:
phæla ‘drop’
banie neoa
‘build (a thing and the
benefit goes to the
doer)’
banie rakha
‘build (a thing keeping
in mind its future use)’
banie phæla
‘build (a thing)’
Effect of the action directed towards the
actor himself (self-beneficiary)
The result of the action is focused
The completion of the action is focused
Table 1: Semantics of a compound verb is the modification or extension of its V1 associate
The meaning of CVs (see column 3)
demonstrates that the core meaning of
banano ‘build’ is preserved in every case.
The verb banano ‘build’, like any other
accomplishment verb, denotes a situation in
which an actor performs some action and the
action has a natural outcome. In case of bari
banano ‘build a house’, for example, the
house is the entity to be built. Which
subpart of V1 semantics do the two light
verbs deoa ‘give’ and neoa ‘take’ profile?
Before we ask that question, let us look at
the semantic requirement of banano ‘build’
in more depth.
Since this is an
accomplishment verb, there exists a
culmination which borne a result of the
action. The semantics of the verb does not
specifically indicate to whom the result of
action is directed. I will call this as
semantics of affectedness. This is based on a
relatively old semantic distinction between
the affected object and effected object
(Fillmore 1968) and Klaiman’s (1991)
analysis of “affective” and “effective” verbs.
The verb banano ‘build’ is unmarked for the
semantics of affectedness. The builder can
build a house for himself (as demonstrated
in (3a)) or he can build a house for the
benefit of a receiver or beneficiary as shown
in (3b):
jonne
3. binu nije-r jonne/ ritu-r
Binu self-gen for Ritu-gen for
bari-ta bana-len
house-cl build-3 hon pt
a. ‘Binu has built the house for himself’
b. ‘Binu built the house for Ritu’
The V2s deoa ‘give’ and neoa ‘take’ remove
this vagueness by categorically focusing on
one particular direction. Light verbs deoa
‘give’ and neoa ‘take’ profile the manner in
which participants are involved in a
situation. The CV banie deoa “build-cp
give” specifies that the effect of the action
directed towards an entity other than the
doer. Ritu is the beneficiary in the following
sentence:
3
4. binu ritu-ke æk-ta bari bani-e di-lo
Binu Ritu-obj one-cl house build-cp give3 pt
‘Binu built a house for Ritu’
The following sentence is awkward which
intends to state that Binu builds the house
for himself and uses the CV banie deoa:
5. ?binu nij-er jonne æk-ta bari
Binu self for one-cl house
bani-e di-lo
build-cp give-3 pt
In such situation, the right choice of V2 is
neoa ‘take’. The CV sequence banie neoa
“build-cp take” profiles the self-directedness
(or self-beneficiary) reading inherent in the
semantics of the verb banano ‘build’ as
shown in (10):
notion of semantics of affectedness that
considers whom or what the result of action
is directed to:
i. Affected type: The result of the action is
directed to the subject. That is, the doer of
the action himself is affected by the action
in the sense that the doer experiences the
result of the action as illustrated in the
following sentence:
7. ami šikkhOk-er kache Oŋko šikh-chi
I teacher-gen near Maths learn-3 pr cnt
‘I am learning Maths from the teacher’
The verb šekha ‘learn’ in (7) entails that the
subject is affected by being benefited from
the action of learning Maths from the
teacher.
Some verbs of this class are given below:
6. binu nij-er jonne æk-ta bari
Binu self-gen for one-cl house
bani-e nieche
build-cp give-3 pt
‘Binu built a house for himself’ (selfbeneficiary)
Many verbs, unlike banano ‘build’, are not
inherently unmarked for the semantics of
affectedness. For example, the meaning of
the verb gOchano ‘foist something on
somebody’ distinctly entails that the result
of the action is directed towards an affected
entity who is not the doer. On the other
hand, verbs such as bhaba ‘think’, khaoa
‘eat’ entails that the doer himself is the
affected entity – the recipient of the result of
the action.
The next section presents semantic type
system for main verbs that select either of
the two light verbs deoa and neoa or both.
3.0 Semantic type of main verbs
In this section we postulate three types of
semantic relations for transitive verbs. They
are i) Affected type, ii) Causally Affected
The
type and iii) Effected type.
classification is primarily motivated by the
khaoa ‘eat’, gela ‘swallow’, neoa ‘take’,
bhaba ‘think’, šekha ‘learn’, paoa ‘get’.
bojha ‘comprehend’, dækha ‘see’,
šona ‘hear’
ii.Causally affected type: The object is a
causally 4 affected entity. That is, unlike
verbs of the previous type, the result of the
action is directed towards a non-doer
entity. For example, the verb oškano
‘instigate’ implies that the subject causes
to raise an emotion in the object who is
affected in the process as exemplified
below:
4
Scholars in their attempt to understand the nature of
‘causation’ have come up with two completely
different viewpoints. They are: (i) bi-event analysis
of causation and (ii) the analysis of causation in terms
of force-dynamic relations among participants. I
use the force-dynamic model of causation. Talmy
(1988) first introduced the concept of force dynamics.
This system represents causation as individuals acting
upon individuals, with some notion of transmission of
force determining which participant is the ‘first’ in the
causal order or the causal chain. Croft (1991) and
Jackendoff (1990) adopt the concept of force-dynamic
relation in interpreting the relation of causation.
4
8. ramu karkhana-r kormi-der hartal kor-te
Ramu factory-gen laborers strike do-inf
oška-cche
instigate-3 pr cont
‘Ramu is instigating the factory laborers
to do strike’
banano ‘build’, kena ‘buy’, khõja ‘search’
rãdha ‘cook’, bhaja ‘fry’, taŋano ‘hang
up’, kata ‘cut (vegetables)’, ãka ‘draw’, ana
‘bring’, kOra ‘do’, kora ‘scrape’, korano
‘collect’, gOra ‘build’, chOka ‘plan’, toka
‘copy’, para ‘bring down’, bacha ‘select’
In this sentence, the ‘laborers’ are the
affected entity because the result of the
action of instigation is directed towards
them. Following are some instances of this
type of verbs:
Following are the prototypical lexical
representation of the semantic types
discussed in this section:
oškano ‘instigate’, kamrano ‘bite’, gõtano
‘thrust’, bOka ‘scold’, põta ‘implant’,
thokrano ‘prick’, mara ‘kill’, ãcrano
‘scratch’, mOckano ‘twist’, chẽra ‘tear’,
bhaŋa (trans) ‘break’, khæpano ‘irritate’,
tatano ‘instigate’, phatano ‘explode’,
janano ‘inform’, porano ‘burn’ bãcano
‘rescue, save’, harano ‘defeat’, jalano
‘inflame, light’, nebano ‘extinguish’,
nabano ‘bring something down’, bholano
‘charm, comfort’, thamano ‘stop, bring to a
halt’, hOtano ‘drive away’, pherano ‘bring
someone back’, dhokano ‘penetrate, cause to
enter’
iii. Effected type: The event associated to
this type entails a logical end-state for the
object 5 . For example, the verb banano
‘build’ entails that the object has come to a
state of existence at the end of activity.
Besides the two regular participants (subject
and object) the semantics of this type of
verbs also presumes the presence of an
affected entity who may not directly
participate in the activity but who might be
benefited from the action that is, the result
of the action is directed towards that entity.
We argued above that the doer himself may
be that affected entity (see the sentence in
(3) for illustration) Unlike verbs of first and
second types, verbs of this type are,
therefore, unmarked for who the results of
the action goes to. Some verbs of effected
type are the following:
5
The idea reflects Levin’s notion of temporal
dependency.
verb class1
E1 = e1:process
EVENTSTR = E2 = e2: state
RESTR = e1 precedes e2
HEAD = e1
ARGSTR =
ARG1 = [1]
ARG2 = [2]
affected type
QUALIA =
FORMAL = affected(e2, [1])
AGENTIVE = Q(e1, [1], [2])
Figure 1. Prototypical lexical presentation
of affected type verbs
verb class2
E1 = e1:process
EVENTSTR = E2 = e2: culmination
RESTR = e1 precedes e2
HEAD = e1
ARGSTR =
ARG1 = [1]
ARG2 = [2]
causally affected type
QUALIA =
FORMAL = affected(e2, [2])
AGENTIVE = Q(e1, [1], [2])
Figure 2. Prototypical lexical structure for
verbs of causally affected type
5
The above two lexical representations
assume the event structure of both types of
verbs to be comprised of two subevents e1
(that denotes a process) and e2 (that refers to
the culmination of the process). The value
of FORMAL of the Qualia structure is,
however, different for the two semantic
types. The lexical structure of verbs of class
1 states that ARG1 who is the doer is the
affected entity; while that of verbs of class 2
shows that the affected entity is the ARG2
by making ARG2 the argument of the
function ‘affected’.
The following lexical representation
presents the semantic structure of verbs of
effected type:
verb class 3
E1 = e1:process
EVENTSTR = E2 = e2: state
RESTR = e1 precedes e2
unify with both light verbs deoa ‘give’ and
neoa ‘take’.
V1 and V2 play a co-operative role in
deciding the semantics of the derived
compound verb. The next section describes
the mechanism by which the co-operation of
the two verbs results in composition of the
semantics of compound verbs. The
mechanism is a kind of co-composition
device (Pustejovsky 1995).
4.0 Semantic Composition of IndoAryan Compound Verbs
The composition of V1 and V2 takes place
basically in terms of FORMAL feature
unification. The unification is licensed by
compatible type specification of V1 and V2.
The type compatibility holds when one of
the following conditions for the QUALIA
structure of V1 and V2, namely QV1 and QV2
respectively, is satisfied:
QV1 and QV2 belong to the same type
ii.QV2 is a subtype of QV1
iii. There exists a unique greatest lower
bound of QV1 and QV2 and that
becomes the QUALIA structure of the
resultant Compound verb.
i.
ARGSTR =
ARG1 = [1]
ARG2 = [2] [physical_obj]
D-ARG = [3] [human]
effected type
QUALIA = FORMAL=result_state(e2, [2])
AGENTIVE = Q(e1, [1], [2])
Figure3: Prototypical lexical structure for
verbs of effected type
As the semantic representation illustrates, I
have postulated a default argument (DARG) in the ARGSTR.
The default
argument is the one who is the recipient or
beneficiary and is not syntactically realized
as an object.
The denotation of this
argument can be the doer of the action
himself and in that case the result of the
action is directed towards the doer. Or it can
be some one other than the doer. Since this
class of verbs is unmarked for to whom the
result of the action is directed to, they can
The light verbs deoa ‘give’ and neoa ‘take’
are taken to be the member of causally
affected type and affected type respectively.
This paper does not discuss how these two
semantic types are logically related to the
semantic types of their full verb
counterparts.
Since the semantic type of QUALIA feature
of the light verb deoa ‘give’ is identical to
that of the main verb of causally affected
type, their unification results in the
following semantic representation of
compound verbs:
6
Causally affected verb_deoa
EVENTSTR =
ARGSTR =
Effected verb_deoa
E1 = e1:process
E2 = e2: culmination
RESTR = e1 precedes e2
E1 = e1:process
EVENTSTR = E2 = e2: state
RESTR = e1 precedes e2
ARG1
= [1]
ARG2 = [2] [physical_obj]
ARGSTR =
causally affected type
QUALIA =
effected typeΛ causally affected
FORMAL = affected(e2, [2])
AGENTIVE = Q(e1, [1], [2])
Figure: 4. Semantic representation
compound verb with V2 deoa ‘give’
ARG1 = [1]
ARG2 = [2] [physical_obj]
ARG3 = [3] [human]
QUALIA =
FORMAL = affected(e1, [3]),
result_state(e2, [2])
AGENTIVE = Q(e1, [1], [2])
of
The Qualia structure of causally affected
verbs and that of their compound variant
with the light verb deoa ‘give’ (as shown in
figure 4) are identical, which implies that the
compound verb does not contain any new
information in terms of the semantics of
affectedness6. However, the co-composition
of the light verb deoa ‘give’ and verbs of
effected type results in a semantic structure
which is very different from that of its
constituent main verb. The unification
produces the semantic representation of a
verbal complex predicate that licenses one
more argument than that of its main verb
counterpart.
The value of D-ARG of
effected type verb becomes the third
argument of the derived compound verb.
What was a default argument for effected
type verbs becomes indirect object for the
compound verbs. The unification adds on a
new function “affected” to the FORMAL
value of the resultant predicate. This is
illustrated in figure 5 below.
Figure 5. The semantic representation of
effected type + deoa ‘give’ compound verb
The unification of the light verb neoa ‘take’
with the effected type verb results in contentsharing between ARG1 and D-ARG. This
implies that the affected entity is the doer
itself.
The FORMAL structure incorporates a
function ‘affected’ that states that the third
argument is affected by the action.
Effected verb_neoa
E1 = e1:process
EVENTSTR = E2 = e2: state
RESTR = e1 precedes e2
ARG1 = [1]
ARGSTR = ARG2 = [2] [physical_obj]
D-ARG = [1]
effected type Λ affected
QUALIA = FORMAL = affected(e2, [1]),
result_state(e2, [2])
AGENTIVE = Q(e1, [1], [2])
6
The enquiry regarding how the causally
affected verb differs semantically from their
corresponding compound with V2 deoa ‘give’ is
beyond the scope of discussion here.
Figure: 6.
7
5.0 Conclusion
In this paper, we proposed a variant of cocomposition mechanism that accounts for
the unification of semantically compatible
main verb and light verb to form a
compound verb in Indo-Aryan languages.
The principle of co-composition as defined
in section 4 also constrains the illegitimate
unification of main verb and light verb. The
difference reflected in the Qualia structure
of compound verb and its component main
verb (see figure 5 and figure 6) maintains
the claim that V2s are not semantically
empty and they contribute to the semantics
of the resultant compound verb.
Reference:
Probal Dasgupta. 1977. The internal grammar of
compound verbs in Bangla. Indian Linguistics,
38(3):68-85.
R. Jackendoff. 1990. Semantic Structures. MIT
Press, Cambridge, MA.
Ronald Langacker. 1987. Foundations of
Cognitive Grammar, 1. Stanford University
Press, Stanford.
S. Paul. 2004. An HPSG Account of Bangla
Compound Verbs with LKB Implementation.
Ph.D dissertation, University of Hyderabad,
Hyderabad.
S. Paul. (2005). The semantics of Bangla
Compound Verbs. Yearbook of South Asian
Languages and Linguistics. 101-112.
William Croft. 1991. Syntactic Categories and
Grammatical Relations: The Cognitive
Organization of Information. University of
Chicago Press, Chicago.
A. Goldberg.
1995. Constructions:
A
Construction Grammar Approach to Argument
Structure. University of Chicago Press,
Chicago.
Alex Alsina. 1996. The Role of Argument
Structure in Grammar. CSLI Publications,
Stanford, CA.
Anvita Abbi. 1991. Semantics of explicator
compound verbs. South Asian Languages,
Language Sciences, 13(2):161-180.
Charles J. Fillmore. 1968. The Case for Case.
Bach and Harms (Ed.): Universals in
Linguistic Theory. New York: Holt, Rinehart,
and Winston,1-88.
Edward W. Hinrichs and T. Nakazawa. 1990.
Subcategorization and VP structure in German.
J. Nerbonne, K. Netter and C. Pollard (eds.),
Proceedings of the Third Symposium on
Germanic Linguistics. Benjamins, Amsterdam.
Gopabandhu Mohanty. 1992. The Compound
Verbs in Oriya. Ph. D. dissertation, Deccan
College Post-Graduate and Research Institute,
Pune.
James Pustejovsky. 1995. The Generative
Lexicon. Cambridge, MA: MIT Press.
M. H. Klaiman. 1981. Volitionality and Subject
in Bengali. Indiana University.
Miriam Butt. 1995. The Structure of Complex
Predicates in Urdu. Doctoral Dissertation,
Stanford University.
Peter Hook. 1974. The Compound Verbs in
Hindi. The Michigan Series in South and
South-east Asian Language and Linguistics.
The University of Michigan.
8