Japanese Dependency Analysis Using the Ancestor-Descendant Relation

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 600–609, Prague, June 2007. c⃝2007 Association for Computational Linguistics Japanese Dependency Analysis Using the Ancestor-Descendant Relation Akihiro Tamura†∗ Hiroya Takamura†† Manabu Okumura†† † Common Platform Software Research Laboratories NEC Corporation a-tamura@ah.jp.nec.com †† Precision and Intelligence Laboratory, Tokyo Institute of Technology, Japan {takamura,oku}@pi.titech.ac.jp Abstract We propose a novel method for Japanese de- pendency analysis, which is usually reduced to the construction of a dependency tree. In deterministic approaches to this task, depen- dency trees are constructed by series of ac- tions of attaching a bunsetsu chunk to one of the nodes in the tree being constructed. Con- ventional techniques select the node based on whether the new bunsetsu chunk and each node in the trees are in a parent-child rela- tion or not. However, tree structures include relations between two nodes other than the parent-child relation. Therefore, we use ancestor-descendant relations in addition to parent-child relations, so that the added re- dundancy helps errors be corrected. Ex- perimental results show that the proposed method achieves higher accuracy. 1 Introduction Japanese dependency analysis has been recognized as one of the basic techniques in Japanese process- ing. A number of techniques have been proposed for years. Japanese dependency is usually repre- sented by the relation between phrasal units called ‘bunsetsu’ chunks, which are the smallest meaning- ful sequences consisting of an independent word and accompanying words (e.g., a noun and a particle). Hereafter, a ‘chunk’ means a bunsetsu chunk in this paper. The relation between two chunks has a di- ∗Akihiro Tamura belonged to Tokyo Institute of Technology when this work was done. Figure 1: Example of a dependency tree rection from the modifier to the modifiee. All de- pendencies in a sentence are represented by a de- pendency tree, where a node indicates a chunk, and node B is the parent of node A when chunk B is the modifiee of chunk A. Figure 1 shows an example of a dependency tree. The task of Japanese dependency analysis is to find the modifiee for each chunk in a sentence. The task is usually regarded as construc- tion of a dependency tree. In primitive approaches, the probabilities of de- pendencies are given by manually constructed rules and the modifiee of each chunk is determined. How- ever, those rule-based approaches have problems in coverage and consistency. Therefore, a number of statistical techniques using machine learning algo- rithms have recently been proposed. In most con- ventional statistical techniques, the probabilities of dependencies between two chunks are learned in the learning phase, and then the modifiee of each chunk is determined using the learned models in the anal- ysis phase. In terms of dependency trees, the parent node of each node is determined based on the likeli- ness of parent-child relations between two nodes. We here take notice of the characteristics of de- pendencies which cannot be captured well only by 600 the parent-child relation. Consider, for …

Content truncated to 500 words