Understanding Unsegmented User Utterances in Real-Time Spoken Dialogue Systems

Understanding Unsegmented User Utterances in Real-Time Spoken Dialogue Systems Mikio Nakano, Noboru Miyazaki, Jun-ichi Hirasawa, Kohji Dohsaka, Takeshi Kawabata* NTT Laboratories 3-1 Morinosato-Wakamiya, Atsugi 243-0198, Japan nakano @ atom.brl.ntt.co.jp, nmiya @ atom.brl.ntt.co.jp, jun @ idea.brl.ntt.co.jp, dohsaka@ atom.brl.ntt.co.jp, kaw @ nttspch.hil.ntt.co.jp Abstract This paper proposes a method for incrementally un- derstanding user utterances whose semantic bound- aries are not known and responding in real time even before boundaries are determined. It is an integrated parsing and discourse processing method that updates the partial result of understanding word by word, enabling responses based on the partial result. This method incrementally finds plausible sequences of utterances that play crucial roles in the task execution of dialogues, and utilizes beam search to deal with the ambiguity of boundaries as well as syntactic and semantic ambiguities. The re- sults of a preliminary experiment demonstrate that this method understands user utterances better than an understanding method that assumes pauses to be semantic boundaries. 1 Introduction Building a real-time, interactive spoken dialogue system has long been a dream of researchers, and the recent progress in hardware technology and speech and language processing technologies is making this dream a reality. It is still hard, however, for com- puters to understand unrestricted human utterances and respond appropriately to them. Considering the current level of speech recognition technology, system-initiative dialogue systems, which prohibit users from speaking unrestrictedly, are preferred (Walker et al., 1998). Nevertheless, we are still pursuing techniques for understanding unrestricted user utterances because, if the accuracy of under- standing can be improved, systems that allow users to speak freely could be developed and these would be more useful than systems that do not. * Current address: N'I"F Laboratories, 1-1 Hikarino-oka, Yoko- suka 239-0847, Japan Most previous spoken dialogue systems (e.g. sys- tems by Allen et al. (1996), Zue et al. (1994) and Peckham (1993)) assume that the user makes one utterance unit in each speech push-to-talk method is used. unit we mean a phrase from representation is derived, and sentence in written language. act in this paper to mean a interval, unless the Here, by utterance which a speech act it corresponds to a We also use speech command that up- dates the hearer's belief state about the speaker's intention and the context of the dialogue. In this paper, a system using this assumption is called an interval-based system. The above assumption no longer holds when no restrictions are placed on the way the user speaks. This is because utterance boundaries (i.e., semantic boundaries) do not always correspond to pauses and techniques based on other acoustic information are not perfect. Utterance boundaries thus cannot be identified prior to parsing, and so the timing of determining parsing results to update the belief state is unclear. On the other hand, responding to a user utterance in real time requires understanding it and updating the belief state in real time; thus, it is impossible to wait for subsequent inputs to determine boundaries. Abandoning full parsing and adopting keyword- based or fragment-based …

Content truncated to 500 words