7th International Conference on Spoken Language Processing
September 16-20, 2002
This paper focuses on syntactic information contained in prosodic features extracted from read Japanese sentences, and describes a method of exploiting it in dependency structure analysis. The basic idea is to make a statistical model of prosodic feature distribution for each dependency distance. Then, by using the Bayes theorem, the dependency distance of each phrase is predicted from a given feature value. A multi-dimensional feature of F0 was effective to improve parsing accuracy, which was sampled from the parabola fitted to the log-F0 contour. It was also shown that the performance was improved more by linearly combining post-phrase pause duration information with the F0 information.
Bibliographic reference. Takagi, Kazuyuki / Kubota, Hajime / Ozeki, Kazuhiko (2002): "Combination of pause and F0 information in dependency analysis of Japanese sentences", In ICSLP-2002, 1173-1176.