A neural network approach to dependency analysis of Japanese sentences using prosodic information

Prosody and syntax are significantly related with each other as has often been observed. In the field of speech synthesis, many efforts have been made to control prosody so that it reflects the syntactic structure of the sentence. However, the inverse problem, recovery of syntactic structure using prosodic information, has not been so much investigated. 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. In this paper, a multilayer perceptron is employed to estimate conditional probability of dependency distance of a phrase given its prosodic feature, i.e., pause duration and F0 contour. Parsing accuracy was improved by combining two different kinds of prosodic information by the perceptron.