Detecting inappropriate clarification requests in spoken dialogue systems

Abstract

Spoken Dialogue Systems ask for clarification when they think they have misunderstood users. Such requests may differ depending on the information the system believes it needs to clarify. However, when the error type or location is misidentified, clarification requests appear confusing or inappropriate. We describe a classifier that identifies inappropriate requests, trained on features extracted from user responses in laboratory studies. This classifier achieves 88.5% accuracy and .885 Fmeasure in detecting such requests.