Example-based Speech Intention Understanding and Its Application to In-Car Spoken Dialogue System

Example-based Speech Intention Understanding and Its Application to In-Car Spoken Dialogue System Shigeki Matsubara†⋆ Shinichi Kimura‡ Nobuo Kawaguchi†⋆ Yukiko Yamaguchi† and Yasuyoshi Inagaki‡ †Information Technology Center, Nagoya University ‡Graduate School of Engineering, Nagoya University ⋆CIAIR, Nagoya University Furo-cho, Chikusa-ku, Nagoya, 464-8601, Japan matubara@itc.nagoya-u.ac.jp Abstract This paper proposes a method of speech inten- tion understanding based on dialogue examples. The method uses a spoken dialogue corpus with intention tags to regard the intention of each in- put utterance as that of the sentence to which it is the most similar in the corpus. The degree of similarity is calculated according to the degree of correspondence in morphemes and dependen- cies between sentences, and it is weighted by the dialogue context information. An exper- iment on inference of utterance intentions us- ing a large-scale in-car spoken dialogue corpus of CIAIR has shown 68.9% accuracy. Further- more, we have developed a prototype system of in-car spoken dialogue processing for a restau- rant retrieval task based on our method, and confirmed the feasiblity of the system. 1 Introduction In order to interact with a user naturally and smoothly, it is necessary for a spoken dialogue system to understand the intentions of utter- ances of the user exactly. As a method of speech intention understanding, Kimura et al. has pro- posed a rule-based approach (Kimura et al., 1998). They have defined 52 kinds of utterance intentions, and constructed rules for inferring the intention from each utterance by taking ac- count of the intentions of the last utterances, a verb, an aspect of the input utterance, and so on. The huge work for constructing the rules, however, cannot help depending on a lot of hands, and it is also difficult to modify the rules. On the other hand, a technique for tagging di- alogue acts has been proposed so far (Araki et al., 2001). For the purpose of concretely deter- mining the operations to be done by the system, the intention to be inferred should be more de- tailed than the level of dialogue act tags such as “yes-no question” and “wh question”. This paper proposes a method of understand- ing speeches intentions based on a lot of dia- logue examples. The method uses the corpus in which the utterance intention has given to each sentence in advance. We have defined the ut- terance intention tags by extending an annota- tion scheme of dialogue act tags, called JDTAG (JDRI, 2000), and arrived at 78 kinds of tags presently. To detail an intention even on the level peculiar to the task enables us to describe the intention linking directly to operations of the system. In the technique for the intention inference, the degree of similarity of each input utter- ance with every sentence in a corpus is calcu- lated. The calculation is based on the degree of morphologic correspondence and that of depen- dency correspondence. Furthermore, the degree of similarity is weighted by using dialogue con- text information. The intention of the utterance to which the maximum …

Content truncated to 500 words