Resolving event noun phrases to their verbal mentions

Abstract

Event Anaphora Resolution is an important task for cascaded event template extraction and other NLP study. Previous study only touched on event pronoun resolution. In this paper, we provide the first systematic study to resolve event noun phrases to their verbal mentions crossing long distances. Our study shows various lexical, syntactic and positional features are needed for event noun phrase resolution and most of them, such as morphology relation, synonym and etc, are different from those features used for conventional noun phrase resolution. Syntactic structural information in the parse tree modeled with tree kernel is combined with the above diverse flat features using a composite kernel, which shows more than 10% F-score improvement over the flat features baseline. In addition, we employed a twin-candidate based model to capture the pair-wise candidate preference knowledge, which further demonstrates a statistically significant improvement. All the above contributes to an encouraging performance of 61.36% F-score on OntoNotes corpus. © 2010 Association for Computational Linguistics.