A twin-candidate based approach for event pronoun resolution using composite kernel

Abstract

Event Anaphora Resolution is an important task for cascaded event template extraction and other NLP study. In this paper, we provide a first systematic study of resolving pronouns to their event verb antecedents for general purpose. First, we explore various positional, lexical and syntactic features useful for the event pronoun resolution. We further explore tree kernel to model structural information embedded in syntactic parses. A composite kernel is then used to combine the above diverse information. In addition, we employed a twin-candidate based preferences learning model to capture the pair wise candidates'preference knowledge. Besides we also look into the incorporation of the negative training instances with anaphoric pronouns whose antecedents are not verbs. Although these negative training instances are not used in previous study on anaphora resolution, our study shows that they are very useful for the final resolution through random sampling strategy. Our experiments demonstrate that it's meaningful to keep certain training data as development data to help SVM select a more accurate hyper plane which provides significant improvement over the default setting with all training data.