Research Of English Event Anaphora Resolution

Anaphora resolution is an important task in natural language processing systems. In the literature, most of previous work on anaphora resolution aimed at entity anaphora which both the anaphor and its antecedent are mentions of the same real world entity. More recently, as the development of topic detection, question answering, event extraction, text understanding, event anaphora resolution becomes more important and essential. This paper analyzes the differences between entity and event anaphora resolution and then focuses on the event anaphora resolution. The study can be concluded as follow.First, we propose an event pronoun resolution framework using a machine learning approach. The instances creation and the features selection are presented in detail. We also explore some methods to balance negative and positive instances. Especially, the methods based on Centering Theory and the distribution of event anaphor and its predicate. The experimental results showed they are very effective.Then, we introduce structural parse features into event anaphora resolution and propose of the strategies of cutting parse tree and analyze their effect on the resolution system. We also explore the effect of semantic features, such as the semantic roles, semantic similarity and their combination. After combining the flat, structural and semantic features our event pronoun resolution system achieves the F-measure of51.67%on the test set. After that, we propose an event noun phrase resolution system based on above machine learning method and the experimental results showed the F-measure achieves41.42%.Finally, English event pronoun detection is a necessary step of the event pronoun resolution and there are no reported research focused on it. We provide a simple approach to detect the English event pronoun and the experimental results show the ACC+and ACC-achieves67.65%and85.23%respectively.