Spoken language understanding method using confidence measure and dialogue history

Abstract

SUMMARYIn the real environment, it is hard for a speech recognizer to avoid misrecognitions completely. However, if misrecognitions occur, user's intentions are usually misunderstood by a conventional language understanding technique, which simply gives priority to the higher rank hypothesis of a speech recognition result (N-best). The utterances in a dialogue are coherent and correct user's intentions might appear in the lower rank hypothesis of N-best. To understand user's speech intentions in the real environment, we propose the language understanding technique that utilizes the dialogue context and confidence measure, which is the word posterior probability. The experimental results show that proposed technique is more efficient (about 15%) than the conventional technique.

Ranking Multiple Dialogue States by Corpus Statistics to Improve Discourse Understanding in Spoken Dialogue Systems

“…Although there have been several attempts to use discourse information for disambiguating speech understanding results [25], [26], the approaches do not allow ambiguities that span over multiple utterances. There is also a body of work that aims to automatically estimate the confidence of slot values using discourse information [15], [27], [28].…”