The use of multiple pause information in dependency structure analysis of spoken Japanese sentences

There is a close relationship between prosody and syntax. In the field of speech synthesis, many investigations have been made to control prosody so that it conforms to the syntactic structure of the sentence. This paper is concerned with the inverse problem: recovery of syntactic structure with the help of prosodic information. In our past investigations, it was observed that duration of inter-phrase pause is most effective among various prosodic features in dependency structure analysis of Japanese sentences. In those studies, only one kind of pause, i.e. the pause that immediately follows a phrase in question was used. In this paper, another kind of pause is employed as a prosodic feature: the pause that immediately follows the succeeding phrase of a phrase in question. It is shown that simultaneous use of the first and second pauses improves the parsing accuracy compared to the case where only the first pause is used.