Japanese Dependency Analysis using Cascaded Chunking

Japanese Dependency Analysis using Cascaded Chunking Taku Kudo and Yuji Matsumoto Graduate School of Information Science, Nara Institute of Science and Technology {taku-ku,matsu}@is.aist-nara.ac.jp Abstract In this paper, we propose a new statistical Japanese dependency parser using a cascaded chunking model. Conventional Japanese statistical depen- dency parsers are mainly based on a probabilistic model, which is not always efficient or scalable. We propose a new method that is simple and efficient, since it parses a sentence deterministically only de- ciding whether the current segment modifies the segment on its immediate right hand side. Experi- ments using the Kyoto University Corpus show that the method outperforms previous systems as well as improves the parsing and training efficiency. 1 Introduction Dependency analysis has been recognized as a basic process in Japanese sentence analysis, and a num- ber of studies have been proposed. Japanese depen- dency structure is usually defined in terms of the relationship between phrasal units called bunsetsu segments (hereafter “segments”). Most of the previous statistical approaches for Japanese dependency analysis (Fujio and Mat- sumoto, 1998; Haruno et al., 1999; Uchimoto et al., 1999; Kanayama et al., 2000; Uchimoto et al., 2000; Kudo and Matsumoto, 2000) are based on a probabilistic model consisting of the following two steps. First, they estimate modification prob- abilities, in other words, how probable one segment tends to modify another. Second the optimal combi- nation of dependencies is searched from the all can- didates dependencies. Such a probabilistic model is not always efficient since it needs to calculate the probabilities for all possible dependencies and cre- ates n˙(n−1)/2 (where n is the number of segments in a sentence) training examples per sentence. In addition, the probabilistic model assumes that each pairs of dependency structure is independent. In this paper, we propose a new Japanese depen- dency parser which is more efficient and simpler than the probabilistic model, yet performs better in training and testing on the Kyoto University Corpus. The method parses a sentence deterministically only deciding whether the current segment modifies seg- ment on its immediate right hand side. Moreover, it does not assume the independence constraint be- tween dependencies 2 A Probabilistic Model This section describes the general formulation of the probabilistic model for parsing which has been ap- plied to Japanese statistical dependency analysis. First of all, we define a sentence as a sequence of segments B = ⟨b1, b2 . . . , bm⟩ and its syntac- tic structure as a sequence of dependency patterns D = ⟨Dep(1), Dep(2), . . . , Dep(m−1)⟩ , where Dep(i) = j means that the segment bi depends on (modifies) segment bj. In this framework, we as- sume that the dependency sequence D satisfies the following two constraints. 1. Japanese is a head-final language. Thus, ex- cept for the rightmost one, each segment mod- ifies exactly one segment among the segments appearing to its right. 2. Dependencies do not cross one another. Statistical dependency analysis is defined as a searching problem for the dependency pattern D …

Content truncated to 500 words