Acquisition and application of contextual role knowledge for coreference resolution.

摘要

Coreference resolution is the process of identifying when two noun phrases (NP) refer to the same entity. This dissertation makes two main contributions to computational coreference resolution.; First, this work contributes a new method for recognizing when an NP is anaphoric. Most pronouns have an antecedent, but many definite noun phrases do not. I present an unsupervised model for learning nonanaphoric definite NPs from a text collection, and I show that it learns lists of these noun phrases with good accuracy. Recall of these NPs increases from 43% to 79%. I also demonstrate that using these lists to filter nonanaphoric definite NPs prior to coreference resolution provides a mechanism for effecting a recall/precision tradeoff. In two distinct testing domains, recall is traded for precision, leading to precision increases from 60% to 73% and from 68% to 82%.; Second, traditional approaches to coreference resolution typically select the most appropriate antecedent by recognizing word similarity, proximity, and agreement in number, gender, and semantic class. This work contributes a new source of evidence that focuses on the roles that an anaphor and antecedent play in particular events or relationships. I show that using contextual role knowledge as part of the coreference resolution process increases the number of anaphors that can be resolved, and I demonstrate an unsupervised method for acquiring contextual role knowledge that does not require an annotated training corpus. A probabilistic model based on the Dempster-Shafer model of evidence is used to incorporate contextual role knowledge with traditional evidence sources. Among the advantages of this model is the capability to assign evidence to a set of candidates when a knowledge source is unable to distinguish among them. In the two testing domains, the F-measure of anaphor/antecedent pairs increases from 0.57 to 0.61 and from 0.57 to 0.63. Recall increases from 46% to 53% and from 42% to 51% with only minor reductions in precision.