Managing ambiguities across utterances in dialogue
David DeVault, Matthew D. Stone · 2007
Task-oriented dialogue systems exploit context to interpret user utterances cor-rectly. When the correct interpretation of a user utterance is ambiguous, a common response is to employ a special process of clarification that delays context update un-til important ambiguities are resolved, so that the main dialogue task can proceed with an unambiguous context. In this pa-per, we describe an implemented dialogue agent which instead translates ambigui-ties in interpretation into uncertainty about which context has resulted from an utter-ance. It then uses question-asking strate-gies, including clarification as a special case of questions about speaker meaning, to manage its uncertainty across multi-utterance subdialogues. We analyze the agent’s use of these strategies in an em-pirical study of task-oriented dialogues be-tween the agent and human users. 1