Spoken dialogue systems are one of the few examples of systems that can intelligently interact with humans to complete a wide variety of tasks. While many spoken dialogue systems have recently been deployed, much research remains to be done on methods to create natural conversational systems easily for new domains. This lecture will first describe two problems with current techniques: (1) The dialogue manager and response generation components of current systems require a massive amount of domain tuning; (2) Current techniques do not support systems that are easily customized to individual users, user groups, or new domains. I will describe several approaches to addressing these problems with methods for automatically learning dialogue and response generation strategies and with techniques for tailoring systems to individual users. I will illustrate these approaches with examples from working spoken dialogue systems.

Marilyn Walker is a Principal Research Scientist at AT&T Labs Research. She completed an M.S. in Computer Science at Stanford University in 1988, and a Ph.D. in Computer and Information Science at the University of Pennsylvania in 1993. Her current research interests include spoken dialogue, models of performance and automatic adaptation for dialogue systems, and multi-modal interfaces. She has published more than 70 journal articles and refereed conference papers on these topics. She co-edited a book entitled Centering Theory in Discourse (Oxford University Press), a special issue of Computational Linguistics on Empirical Methods in Discourse, and a special issue of Computer Speech and Language on Spoken Language Generation. She is current Chair of the Evaluation Committee for the DARPA Communicator program, focusing on methods for learning performance models for spoken dialogue systems from dialogue corpora.