On the Utility of Decision-Theoretic Hidden Subdialog

A spoken dialog system typically characterizes a domain task with multiple states interconnected by actions or thresholds as transitions between states. As the system attempts to solicit a piece of information from the user, it may have to engage in a hidden subdialog, or error handling within a particular state, before transitioning to a new state. Hidden subdialogs generally center on illocutionary repairs such as a request for repetition or confirmation of a heard utterance. We summarize what we believe to be the distinct advantages of representing error handling in a hidden subdialog as decision making under uncertainty. We motivate the discussion with examples drawn from dialog systems built within the Conversational Architectures Project at Microsoft Research.