Probabilistic Resolution of Anaphoric Reference

This paper describes the use of a Bayesian network to resolve anaphora by probabilistically combining linguistic evidence. By adopting a Bayesian approach, we are able to combine diverse evidence in a principled way, extend current understanding of linguistic phenomena by quantifying relationships empirically, and better model the non-deterministic role of linguistic evidence in resolution of anaphora. We also briefly discuss our current research into the use of learning techniques to automatically construct Bayesian networks from data.