Modelling Human Clarification Strategies

We model human responses to speech recognition errors from a corpus of human clarification strategies. We employ learning techniques to study 1) the decision to either stop and ask a clarification question or to continue the dialogue without clarification, and 2) the decision to ask a targeted clarification question or a more generic question. Targeted clarification questions focus specifically on the part of an utterance that is misrecognized, in contrast with generic requests to ‘please repeat’ or ‘please rephrase’. Our goal is to generate targeted clarification strategies for handling errors in spoken dialogue systems, when appropriate. Our experiments show that linguistic features, in particular the inferred part-ofspeech of a misrecognized word are predictive of human clarification decisions. A combination of linguistic features predicts a user’s decision to continue or stop a dialogue with accuracy of 72.8% over a majority baseline accuracy of 59.1%. The same set of features predict the decision to ask a targeted question with accuracy of 74.6% compared with the majority baseline of 71.8%.1