Deeplistener: harnessing expected utility to guide clarification dialog in spoken language systems
Eric Horvitz, Tim Paek · 2000
We describe research on endowing spoken language systems with the ability to consider the cost of misrecognition, and using that knowledge to guide clarification dialog about a user’s intentions. Our approach relies on coupling utility-directed policies for dialog with the ongoing Bayesian fusion of evidence obtained from multiple utterances recognized during an interaction. After describing the methodology, we review the operation of a prototype system called DeepListener. DeepListener considers evidence gathered about utterances over time to make decisions about the optimal dialog strategy or realworld action to take given uncertainties about a user’s intentions and the costs and benefits of different outcomes.