A Framework for Soliciting Clarification from Users During Plan Recognition

In previous work, we used plan recognition to improve responses to users in an advice-giving setting. We then characterized when it was worthwhile to engage a user in clarification dialogue--the cases where plan ambiguity mattered to the formulation of a response. Our current research develops detailed algorithms for selecting what to say during the clarification dialogue. We propose a default strategy for selecting a clarifying question, together with a variety of options to reduce the length of the clarification dialogue. Each clarifying question is introduced in order to prune the set of possible plans. But the system will never overcommit in recognizing the user's plan and thus will never have to backtrack into a debugging dialogue with the user. In all, we now have a more precise formulation for what to say during clarification dialogues, and valuable characterizations of decisions which an advice-giving system must make in generating a dialogue with a user.