Ranking Help Message Candidates Based on Robust Grammar Verification Results and Utterance History in Spoken Dialogue Systems

Kazunori Komatani, Satoshi Ikeda, Yuichiro Fukubayashi, Tetsuya Ogata and Hiroshi G. Okuno

SIGDIAL Workshop on Discourse and Dialogue (SIGDIAL 2009)
Queen Mary University of London, September 11-12, 2009


We address an issue of out-of-grammar (OOG) utterances in spoken dialogue systems by generating help messages for novice users. Help generation for OOG utterances is a challenging problem because language understanding (LU) results based on automatic speech recognition (ASR) results for such utterances are always erroneous as important words are often misrecognized or missed from such utterances. We first develop grammar verification for OOG utterances on the basis of a Weighted Finite-State Transducer (WFST). It robustly identifies a grammar rule that a user intends to utter, even when some important words are missed from the ASR result. We then adopt a ranking algorithm, RankBoost, whose features include the grammar verification results and the utterance history representing the user’s experience.