Speech Prosody 2004
This paper deals with the problem of recovering syntactic structures of sentences by using the prosodic information extracted from spoken versions of the sentences. Prosodic information has proven to be effective to disambiguate syntactic structures, which is not utilized in a conventional rule-based parser. In our previous works, the duration of pauses at phrase boundaries has been found to be consistently and dominantly effective for improving parsing accuracy of read Japanese sentences, although the evaluation was limited to a small set of test speakers. In this paper, dependency analysis using pause information was conducted in a speaker-independent manner by using larger amount of speech data read by 316 speakers. Effects of pause duration normalization were observed, although the parsing accuracy was lower than that in speaker-dependent case.
Bibliographic reference. Takagi, Kazuyuki / Ozeki, Kazuhiko (2004): "Dependency analysis of read Japanese sentences using pause information: a speaker independent case", In SP-2004, 595-598.