We understand natural language discourse so well because we know so much. We are able to draw the inferences necessary to tie the various parts of a discourse together, and how we do this is perhaps the central problem in natural language understanding. In this talk, I will first show that many representational problems can be bypassed by reifying states and events, resulting in a very simple picture of compositional semantics. Then I will show how abduction, or finding the best explanation for the content of a text, solves a wide range of pragmatics problems, including coreference resolution, the interpretation of metonymy and metaphor, and discovering discourse structure. Finally, I will discuss an effort to build an adequate knowledge base for natural language understanding, by both manual and automatic means, in two areas -- the structure of events and goal-directed behavior.
To view a recording of this talk click here (You will need a Rutgers NetID and password)