Automatic Lexical Acquisition Based on Statistical Distributions

We automatically classify verbs into lexical semantic classes, based on distributions of indicators of verb alternations, extracted from a very large annotated corpus. We address a problem which is particularly di cult because the verb classes, although semantically di erent, show similar surface syntactic behavior. Five grammatical features are su cient to reduce error rate by more than 50% over chance: we achieve almost 70% accuracy in a task whose baseline performance is 34%, and whose expert-based upper bound we calculated at 86.5%. We conclude that corpus-driven extraction of grammatical features is a promising methodology for ne-grained verb classi cation.