Japanese Dependency Structure Analysis Based on Maximum Entropy Models

Proceedings of EACL '99 Japanese Dependency Structure Analysis Based on Maximum Entropy Models Kiyotaka Uchimoto t Satoshi Sekine$ Hitoshi Isahara t tCommunications Research Laboratory Ministry of Posts and Telecommunications 588-2, Iwaoka, Iwaoka-cho, Nishi-ku Kobe, Hyogo, 651-2401, Japan [uchimot o i isahara] ©crl. go. j p SNew York University 715 Broadway, 7th floor New York, NY 10003, USA sekine~cs, nyu. edu Abstract This paper describes a dependency structure analysis of Japanese sentences based on the maximum entropy mod- els. Our model is created by learning the weights of some features from a train- ing corpus to predict the dependency be- tween bunsetsus or phrasal units. The dependency accuracy of our system is 87.2% using the Kyoto University cor- pus. We discuss the contribution of each feature set and the relationship between the number of training data and the ac- curacy. 1 Introduction Dependency structure analysis is one of the ba- sic techniques in Japanese sentence analysis. The Japanese dependency structure is usually repre- sented by the relationship between phrasal units called 'bunsetsu.' The analysis has two concep- tual steps. In the first step, a dependency matrix is prepared. Each element of the matrix repre- sents how likely one bunsetsu is to depend on the other. In the second step, an optimal set of de- pendencies for the entire sentence is found. In this paper, we will mainly discuss the first step, a model for estimating dependency likelihood. So far there have been two different approaches to estimating the dependency likelihood, One is the rule-based approach, in which the rules are created by experts and likelihoods are calculated by some means, including semiautomatic corpus- based methods but also by manual assignment of scores for rules. However, hand-crafted rules have the following problems. • They have a problem with their coverage. Be- cause there are many features to find correct dependencies, it is difficult to find them man- ually. • They also have a problem with their consis- tency, since many of the features compete with each other and humans cannot create consistent rules or assign consistent scores. • As syntactic characteristics differ across dif- ferent domains, the rules have to be changed when the target domain changes. It is costly to create a new hand-made rule for each do- main. At/other approach is a fully automatic corpus- based approach. This approach has the poten- tial to overcome the problems of the rule-based approach. It automatically learns the likelihoods of dependencies from a tagged corpus and calcu- lates the best dependencies for an input sentence. We take this approach. This approach is taken by some other systems (Collins, 1996; Fujio and Mat- sumoto, 1998; Haruno et ah, 1998). The parser proposed by Ratnaparkhi (Ratnaparkhi, 1997) is considered to be one of the most accurate parsers in English. Its probability estimation is based on the maximum entropy models. We also use the maximum entropy model. This model learns the weights of given features from a training corpus. The weights are calculated based on the …

Content truncated to 500 words