Empirical evaluations of pronoun resolution

This thesis describes research into ways of improving pronoun resolution algorithms. Many systems today perform well, as high as 80% accuracy in some cases over news article domains, but resolving the other 20% requires additional complex information that isn't usually available to systems. In this project we investigate whether syntactic ranking preferences, discourse structure, and lexical semantics can improve an existing pronoun resolution system to close the “20%” gap. Our results show that syntactic ranking preferences and lexical semantics can boost performance in newspaper and spoken dialogue domains.