Semantics, Dialogue, and Reference Resolution

摘要

Most pronoun resolution research has focused on written corpora while using syntactical and surface cues. Though big gains have been made in this domain with those methods, it is difficult to do better than the 80% coverage in these domains without some world or semantic knowledge. We investigate this issue by incorporating rich semantic information into a proven reference resolution model over a very difficult domain of human-human task- oriented dialogues. Our results show that semantic information greatly improves performance and can even be viewed as a substitution for the usual syntactic filters.