Clarification Questions with Feedback

In this paper, we investigate how people construct clarification questions. Our goal is to develop similar strategies for handling errors in automatic spoken dialogue systems in order to make error recovery strategies more efficient. Using a crowd-sourcing tool [7], we collect a dataset of user responses to clarification questions when presented with sentences in which some words are missing. We find that, in over 60% of cases, users choose to continue the conversation without asking a clarification question. However, when users do ask a question, our findings support earlier research showing that users are more likely to ask a targeted clarification question than a generic question. Using the dataset we have collected, we are exploring machine learning approaches for determining which system responses are most appropriate in different contexts and developing strategies for constructing clarification questions