Cognitive Belief Modeling for Naturalistic Dialog Management
Students of human conversation observe that people naturally and automatically build mental models of each others' intents and understanding in real time, as we hear and respond to speech. These mental models drive social devices of conversational dialog management, including turn control, affirmation, and repair of breakdowns in shared knowledge. Can we give AI assistants these same capabilities? This talk will illustrate how a cognitive belief model can support naturalistic back-and-forth communication of structured data. In the illustrative example agent, the dialog manager is hand-crafted to hold probabilistic beliefs. This raises the question of how much a head start rule-based dialog managers have over learning-based approaches.
Eric Saund is a research scientist, developer, and consultant whose work spans multiple fields of intelligent systems, including conversational agents, document analytics, perceptually supported user interfaces, machine learning, and computational vision. Eric received a B.S. from Caltech and a Ph.D. in Cognitive Science from MIT. He served in research and research management roles at the Xerox Palo Alto Research Center, he has published widely, his prototypes have served thousands of users, and he has been awarded over 50 patents to date.