Semantic Classification of Verbs Based Statistical Syntactic
Features
Suzanne
Stevenson
Department of Computer Science
University of Toronto
Mittwoch, 26.09.2001, 14 Uhr c.t., D6-135
A semantic classification of verbs can be useful in the organization
of lexical information, but is time-consuming and difficult to produce
manually. One important level of classification of verbs is that of
predicate-argument structure -- how an action or state is related to
its participants (i.e., who did what to whom). In this work, we
describe machine learning experiments to automatically classify three
major types of English verbs, based on their predicate-argument
structure -- specifically, the conceptual roles they assign to
participants. These conceptual roles are the way in which the
relational semantics of the verb is represented at the syntactic
level, and thus serve as a link between syntax and semantics. Our
hypothesis then is that carefully selected syntactic features gleaned
from the use of a verb in a corpus may help in inducing the underlying
semantic classification of the verb.
We use linguistically-motivated statistical indicators extracted from
a large annotated corpus to train a classifier, achieving 69.8%
accuracy for a task whose baseline is 34%. Our results validate our
hypotheses that knowledge about predicate-argument relations is useful
in semantic verb classification, and that it can be gleaned from a
corpus by automatic means. We thus demonstrate an effective
combination of deeper linguistic knowledge with the robustness and
scalability of statistical techniques.