Handling Failures in Human-Computer Conversation.pdf

Handling Failures in Human-Computer Conversation Tim Paek Department of Psychology Stanford University Stanford, CA 94305 U.S.A. paek@psych.stanford.edu Abstract. Uncertainty plays a central role in the handling of failures in conversation. People manage uncertainty by establishing and maintaining the mutual belief that their utterances have been understood well enough for current purposes, a process referred to as grounding. In this paper, we propose a computational architecture for supporting grounding in human-computer spoken dialog. We delineate four levels of analysis and describe representations and inference procedures for resolving key uncertainties that may lead to communication failures. We illustrate the architecture by reviewing inter- actions between a user and a prototype spoken dialog system called the Bayesian Re- ceptionist. Keywords: Bayesian Receptionist, Conversation, Decision Theory, Human-Computer Interaction, Misunderstanding, Spoken Dialog 1. Introduction When people engage in a conversation, they typically do so with the intent of making themselves understood. To do this they need to make sure, as they speak, that the other participants are at the same time attending to, hearing, and understanding what they are saying. Since uncertainty about these aspects can result in communication failures, people collaborate to establish and maintain the mutual belief that their utterances have been understood well enough for current purposes (Clark and Schaefer, 1987, 1989). An example of this process is feedback. To establish the mutual belief that an utterance was sufficiently understood, people will give feedback of their understanding through head nods or acknowledgements, such as “uh huh.” If they do not understand, they will try to clear up their uncertainties. In short, people coordinate not only what they say but also their beliefs about what they mutually understand. Researchers in linguistics, psychology, and artificial intelligence have argue