Sentence verification in spoken dialogue system

In spoken dialogue systems, sentence verification technique is very useful to avoid misunderstanding user’s intention by rejecting out-of-domain or bad quality utterances. However, compared with word verification and concept verification, sentence verification has been seldom touched in the past. In this paper, we propose a sentence verification approach which uses discriminative features extracted from the edit operation sequence. Since the edit operation sequence indicates what kinds of errors (i.e., insertion, deletion and substitution errors) may occur in the hypothetical concept sequence, it conveys sentence-level information for evaluating the quality of system’s interpretation for the user’s utterance. In addition, a sentence verification criterion concerning precision and recall rates of hypothetical concepts is also proposed to pursue efficient and correct spoken dialogue interactions. Compared with the verification method using acoustic confidence measure, the proposed approach reduces 17.3% of errors.