Analysis of Japanese Sentence Structure Based on the Sentence Length Distribution and Dependency Relations among Segments

Abstract

In previous studies about Japanese sentence length, two different types of the model for sentence generation have been proposed by Sasaki: one is a multiplicative stochastic model resulting in log-normal sentence length distribution and the other is an additive stochastic model resulting in the negative binomial sentence length distribution.
  In the present study, motivated by Sasaki's suggestion, we examined the structure of dependency trees and checked whether those models could explain the obtained structure of dependency trees. To do that, we used Kyoto University Text Corpus (33,082 sentences) which includes the information of dependency relations among segments.
  As a result, we found that the structure of the dependency trees did not accord with the expectation of multiplicative nor additive stochastic process.