Utterances in situated activityare about the world.  At the same time, speakers use real-worldactivity along with speech and gesture to help make theirpoints.  To account for these integrated contributions, wesuggest an approach to discourse coherence which can distinguishbetween the concrete, useful changes thatpractical actions bring about in the world and the distinctive contributionsthose actions may also signal in a conversation.  In particular, we argue that coherence relations must sometimes link practical actions to ongoingdiscourse, because interlocutors use those actions to coordinateupdates to the conversational record.Conversely, coherence relations

must sometimes link utterances to aspects of the nonlinguisticcontext, to capture the context-dependent nature of theirinterpretation.  Our approach promises better empiricalcoverage and more straightforward system building.

Joint work with Una Stojnic andErnie Lepore (Rutgers).

Matthew Stone is AssociateProfessor in the Computer Science Department and the Center for CognitiveScience at Rutgers. He got his PhD in 1998 from the University of Pennsylvania.He studies computational models of conversation, particularly models ofutterance production, for intelligent agents that interact naturally with humanpartners. He recently concluded a term on the editorial board of the journalComputational Linguistics and served as program co-chair for the 2007 NorthAmerican Association for Computational Linguistics Human Language TechnologyConference (NAACL HLT).