Syntactic Features Based Pronoun Resolution

Coreference largely exists in discourse and dialogue in natural languages. For human understanding, it makes sentences more concise and makes the topic more brilliant. However, it makes languages more difficult to process computationally. With the increasing demand of discourse processing, coreference resolution plays more and more important role in different fields such as machine translation, information extraction, automatic summarization, question and answer system.For pronoun resolution, a parse tree that covers a pronoun and its antecedent candidate could provide us much syntactic information related to the pair. However, because of the complex structures and the difficulties to represent it effectively, traditionally, these features such as syntactic roles and binding constrains are calculated by mining the parse trees, and then could be used for resolution by using manually and heuristicly designed rules. These methods have some limitations, such as: dependent to the researchers'linguistic knowledge too much; hard to make use of implicit information in parse trees; lack of relations with other elements in context.In this paper, we focus on how to represent parse structures effectively and how to find useful information in context by making use of parsing. We propose 3 approaches:First order logic could express the relations between terms. We make use of this feature to represent the parse tree path between the pronoun and antecedent candidate as a set of predicates and apply Inductive Logic Programming which could learn hypothesis directly from background knowledge.Make use of the structured syntactic knowledge to do pronoun resolution. We propose a kernel-based method to incorporate the information from parse trees together with other normal features to learn a classifier and do resolution..Dependency grammar benefits to represent the relation between words. We propose a dependency parsing based method for Chinese pronoun resolution, designing effective syntactic role features and word sense similarity and word relevance features in related to the dependent words. Our contribution lies in constructing structured syntactic features, by making use of dependency parsing, finding useful information in context for pronoun resolution.