Abstract: Spoken dialogue systems are computer systems that converse with humans using spoken natural language. They have traditionally been built under the assumption that the human user's utterances are mutually understood when they are first spoken unless there is evidence to the contrary. However, studies of dialogues between humans have shown that humans provide a broad range of both positive and negative evidence to ensure that information is appropriately added to their common ground.

This research studies how that evidence can be formalized and used in spoken dialogue systems. A dialogue system for military training in an immersive virtual environment is implemented and a corpus of human interactions is analyzed to produce a model of evidence of understanding. Patterns in the presentation of this evidence are used to identify the degrees of groundedness that material passes through before becoming sufficiently grounded. Algorithms to identify the evidence of understanding and the degrees of groundedness are implemented and evaluated. Algorithms for dialogue management using this model are implemented in a virtual human in a military questioning domain.

Short Bio: Antonio Roque is a CS PhD candidate at the University of Southern California, completing research in dialogue modeling for spoken dialogue systems at USC's Institute for Creative Technologies. From 2000 - 2003 he worked on a natural language intelligent tutoring system at the University of Pittsburgh; prior to that he worked as a programmer and computer consultant for academic and business customers.