A Twin-Candidate Model for Learning Based Coreference Resolution

Abstract

The purpose of the thesis is to find an effective learning model for coreference resolution. The traditional single-candidate model only considers one individual antecedent candidate at a time for its learning, and thus cannot capture the preference relationship between competing candidates. To overcome this limitation, the thesis proposes a twin-candidate learning model which recasts antecedent selection as a preference classification problem. The model learns a classifier that can determine the preference between two competing antecedent candidates of an anaphor, and then choose the antecedent based on the ranking of the candidates. The thesis explores how to use the twin-candidate model to identify the antecedent from the candidates of an anaphor. Further it investigates how to deploy this model to coreference resolution. Also it discusses how to represent the preference knowledge in the model. The experimental results indicate that the model is effective for both anaphora resolution and coreference resolution.