E ectiveness of Prosodic Information in Dependency Analysis of Japanese Sentences

This paper is concerned with measuring the amount of syntactic information contained in prosodic features of read Japanese sentences. Five prosodic features were chosen, and statistical relation between those features and interphrase dependency distance was estimated from a speech database. Then a number of experiments on dependency analysis of Japanese sentences were conducted with the minimum total penalty method by using the statistical relation. The result showed a signi cant improvement in parsing accuracy and dependency accuracy compared with a baseline result involving no prosodic information. The duration of pause between adjacent phrases was slightly more e ective than other prosodic features such as pitch and power in estimating the syntactic structure of read Japanese sentences.