Defining Verb Semantic Classes for French and Their Semantic Characterization

Abstract
In this document, we first show how verb semantic
classes can be formed from syntactic descriptions
(called contexts) realized in a way close to B.
Levin’s work. The verb class system we obtain for
French is radically different from the system obtained for English. We explain why and what kind
of additional criteria (thematic grids) should be
used to get a class system which has the good
properties of the English one. The different softwares that permit this classification are presented.
Finally, we show that verb semantic classes is a
powerful means for organizing the various semantic information related to verbs, useful in NLP
applications. We show how various, coherent semantic or conceptual systems can be associated
with these verb classes: selectional restrictions,
thematic grids, aspectuality, and the Lexical Conceptual Structures (LCS).
1. Introduction
This work is primarily based on B. Levin’s work
(Levin 93) for English, where she shows that the
syntactic behavior of verbs is essentially predictable from some aspects of their semantics. By
syntactic behavior, she means the way arguments
are realized in a syntactic form with respect to the
predicate, how they can move (e.g. to define ergative or passive forms) and when they can be deleted. These movements and deletions are called
alternations. It is important to note that the operations described by amternations are strictly lexically-based (therefore they do not include Wh- or
NP-movements).
Although B. Levin’s work corresponds to a certain
level of granularity in the linguistic description
which has a certain degree of stability, we do not
think that it really introduces a new level with a
theoretical status in syntax or in semantics. Rather,
we consider this work as a very useful, practical
and relatively comprehensive one, which can form
a good perspective and a good practical basis for
the extraction and for the organization of lexical
data.
We have substantially reformulated B. Levin’s
system in order to avoid a number of difficulties
inherent to her approach, which do not seem to us
to be central to her system (as she recognizes it in
recent discussions), and to make the system more
declarative and more usable in NLP systems. Instead of considering the alternation system of B.
Levin, we have reformulated this notion into a
more declarative one: the context system. Then,
we can assign to each verb-sense the set of syntactic contexts it accepts. By syntactic context we
mean a set of distribution frames (a frame may
include several distributions) where the category
and some additional syntactic and semantic information is used to describe the nature and the form
of the possible complements and the subject of a
verb. This is presented in (Saint-Dizier and Marrafa 96) (and also at the last ACH-ALLC conference). We now show how this system is used to
construct semantic verb classes and how semantic
information can be associated with these classes.
2. Verb classes
Verb classes are formed out from verbs having
similar sets of contexts by specific software in
Prolog that we have defined. For about 1700 usual
verbs of French, we obtain about 400 verb classes
(This is not a surprising result). A verb with different senses appears in different classes. We, in
fact, consider that a class could define the ‘boundaries’ of a verb-sense and could therefore introduce a ‘data-driven’ way of defining polysemy for
verbs.
Classes produced are very different from the English ones. We have, on the one hand, very large
classes of verbs, where verbs are not necessarily
semantically related (subsets of semantically related verbs are observed), and on the other hand,
very small verb classes, with verbs having very
close meanings. We also get classes with different
types of verbs, often accepting few contexts. In
general the more contexts a class is based on, the
better its quality is. Good results are obtained with
classes associated with at least 5 contexts. The
global semantic relatedness is about 61%, which
is not very good. This rate has been computed
from a WordNet classification (Fellbaum 90) that
we have realized following the principles of
WordNet for English and using exactly the same
verb-senses. This is due to the fact that in French
we have less contexts than there are alternations
in English. It is therefore not surprising that they
form a less dicriminatory system w.r.t. verb class
formation. Also, our contexts do not contain semantic information (as those of B. Levin’s do).
We have realized a detailed analysis of the results
in order to be able to improve the classification
system, so that we can obtain, on an as automatic
as possible way, semantic verb classes of a quality
equivalent to those obtained by B. Levin for English. We will consider 2 principles: the taking
into account of a few exceptions (which slightly
improves the quality of semantic classes), and the
introduction, step by step, and on a global basis of
very general purpose semantic criteria.
We have first conducted a second experiment
where, similarly to B. Levin, we allow exceptions
in classes. To be more precise, we allow verbs in
a given class to have one context more or less than
the set of contexts established as the norm of the
class. This flexibility can be adjusted to only those
contexts estimated to be less important. Besides
allowing the grouping of verbs which have very
close syntactic behaviors, this flexibility also allows us to weaken the effect of possible judgement
errors in the construction of the list of contexts
associated with every verb (in some cases, it is not
very easy to say if a verb definetely accepts a
context or if it is just a marginal use). In this case,
we obtain slightly better results, in particular for
the smaller classes which are now larger and have
a size which can really be considered as that of a
class. Larger classes have not changed much. We
now need to elaborate criteria for decomposing
these classes into smaller units where verbs are
semantically related.
There are two basic sources of information which
can be considered to decompose these latter classes: selectional restrictions and theta-grids. The
most reliable and relevant system are theta-grids,
for which we have defined relatively fine-grained
theta-roles (Pugeault et al. 94), inspired from
(Dowty 91). We have decomposed these larger
classes into smaller units based on the contents of
thematic grids: e.g. the presence of a localization
(spatial, temporal or abstract), of a means or of an
incremental beneficiary theme, etc. We then obtain a set of verb semantic classes comparable to
the results obtained for English by B. Levin. Notice that thematic grids alone (i.e. without contexts) would not have been sufficiently discriminatory to construct verb semantic classes, since
there are hundreds of verbs with e.g. a volitional
agent and a general theme).
We have extended our classification system in
order to take into account these basic semantic
criteria. We then get much better classes, which
can be hierarchically organized (an evaluation has
been carried out and is given in (Saint-Dizier
95b).) A global result of semantic relatedness of
89% is obtained. There still remains some work to
be done by hand to go beyond this rate, in order to
deal with exceptional verb behaviors and with
verbs which have been misplaced for various reasons. We have identified a few methods for these
manual tasks, so that they can be reproduced for
other subsets of verbs, of general purpose and
related to technical domains, and also for verbs of
different languages.
3. Semantic characterization
The last point of our research is to show that the
verb semantic class system is a very powerful
system for organizing lexical semantics data, besides the well-known essential lexical semantics
relations such as isa, part-of, synonymy and oppositions. These relations are essentially paradigmatic whereas the relation introduced by verb classes
is essentially analytic.
In a first stage, we have constructed verb semantic
classes that can be grouped into families such as:
the ‘say’, ‘manage’, ‘transfer of possession’, ‘movement’ and ‘psychological’ verb families (SaintDizier 95). These families are more general than
those introduced in B. Levin’s classification. With
our verb sample, we have identified so far 16 verb
families. These families can be further decomposed into sub-families according to precise criteria
related to the semantics of the class. For example,
we have the sub-family ‘transfer of possession
with retro-transfer of money’ a subclass of the
family of transfer of possession. A subfamily contains several verb semantic classes. This decomposition introduces an isa hierarchy, represented
as a directed graph. We can then associate with
every family a set of semantic information (such
as selectional restrictions, fragments of thematic
grids and LCS-based primitives) which is inherited by the subfamilies and the verb semantic classes. Then, at the level of each subfamily (and next,
verb semantic class), we can associated more precise or additional semantic information. This is not
an easy task: it requires the introduction of various
semantic elements related to different ontological
dimensions. So far, we have sucessfully decomposed and described the family of transfer of possession (the ‘say’ family is in progress). This work
has then been confirmed in most of its aspects (and
extended, e.g. to identify prototypical verbs) by
psychological experiments carried out on a quite
large population of children (10 years old) and
adults (students) (Dubois and Saint-Dizier 96).
The introduction of fragments of LCS representations (represented by typed-lambda expressions where underspecified elements are in the
scope of a lambda-expression) is of much interest.
First it shows us how a verb sense can be constructed from more generic and abstract senses specified at the level of families and subfamilies. Next,
it introduces methodological elements to define
additional primitives, central to a family, in the
spirit of those already defined in the LCS. Finally,
it shows the limits of the LCS and opens research
directions on a concrete basis.
241
4. Applications
In terms of applications, the first point to mention
is the classification programmes we have defined
and the methods defined around them, their use in
different situations and languages, and the methods to use for the tasks which cannot be done on
an automatic basis.
The second point is the creation of a quite large
lexicon for French of predicative forms with a
graphic interface that allows users to navigate in
the semantic verb classes and to have access to the
different semantic forms. This lexicon is associated with parametrized procedures that allow the
construction of a sublexicon from that lexicon,
specifically designed for an application. This lexicon can also be used for teaching purposes.
The last point is the use of this lexicon in information retrieval applications where we want to be
able to extract structured information (predicates
and arguments) and to be able to represent it in a
generic way (using LCS, for example, combined
with a few ontological elements). Such an application has been successfully carried out (Pugeault
et al. 94) for the French National Electricity company) and has been evaluated in depth.
Bibliography (Short)
Dorr, B., Machine Translation: a view from the
lexicon, MIT Press, 1993.
Dowty, D., Thematic Proto-roles and Argument
Selection, Language, vol. 67-3, 1991.
Fellbaum, C., The English Verb Lexicon as a
Semantic Net, International Journal of Lexicography, 3/4, 278-301, 1990.
Grimshaw, J., Argument Structure, MIT Press,
1990.
Jackendoff, R., The Status of Thematic Relations
in Linguistic Theory, Linguistic Inquiry 18,
369-411, 1987.
Jackendoff, R., Semantic Structures, MIT Press,
1990.
Levin, B., English Verb Classes and Alternations,
the University of Chicago Press, 1993.
Napoli, D.J., Predication Theory, Cambridge University Press, 1990.
Pugeault, F., Saint-Dizier, P., Monteil, M.G.,
Knowledge Extraction from Texts: a method
for extracting predicate-argument structures in
texts, in proc. Coling 94, Kyoto, 1994.
Saint-Dizier, P., Viegas, E., (Eds.), Computational
Lexical Semantics, Cambridge University
Press, 1995.
Saint-Dizier, P., A semantic classification of
French verbs based on B. Levin’s approach,
research report, IRIT, 1995b.
Saint-Dizier, P., Marrafa, P., Constructing a
knowledge base for describing the general semantics of verbs, in vol. 6 of ‘Research in
Computing for the Humanities’, Oxford Univ.
Press, 1996.
Speas, M.J., Phrase Structure in Natural Language, Kluwer Academic Press, 1990