Pause information for dependency analysis of read Japanese sentences

The work presented in this paper is devoted to the modeling of distribution of post-phrase pause duration for dependency analysis of read Japanese sentences. The pause information is incorporated into a dependency structure parser, which handles numerical information as part of linguistic knowledge. A series of our previous works has shown that the duration of pauses is constantly e ective to improve the parsing accuracy among other prosodic features. This paper aims to improve the parsing accuracy by reforming pause distribution functions to better t the actual distribution of pause duration. First, the pause distribution is divided into two parts, each of which is represented by a separate model: one by a discrete probability model, and the other by a bimodal p.d.f. Secondly, pause duration is calculated in log scale. All the experiments show that these modi cations improve parsing accuracy.