Forest Driven Dependency Analysis Enhanced by Japanese Clause Structure Estimation

Forest Driven Dependency Analysis Enhanced by Japanese Clause Structure Estimation Satoshi Kamatani, Kentaro Furihata, and Tetsuro Chino Corporate Research & Development Center, Toshiba Corporation 1, Komukai Toshiba-cho, Saiwai-ku, Kawasaki, 212-8582, Japan e-mail: satoshi.kamatani@toshiba.co.jp Abstract. For applications of the natural language processing, such as spoken translation, it is the most important characteristic to accept various inputs. We analyze a speech input as a sequence of speech fragment and achieve high acceptability. But it is difficult to determine whether each fragment/unit should be interpreted separately or not. This sort of problem is shared by several applications. We regard a clause and dependency structures as one of the clue to solve this type of problem. Because, even in a spoken language domain, a clause represents a syntactic unit and a dependency represents a semantic unit. In this paper, we describe the dependency analysis method on the syntax forest structure enhanced by Japanese clause structure estimation. It also contributes to improve on translation in manual evaluation from 77.8% up to 78.8%. 1 Introduction It is the most important characteristic for applications of the natural language processing, such as a spoken language translation (SLT), to accept various inputs. Its capability is directly reflected on availability. We design a SLT system based on this concept. Our SLT system can be divided into two parts. The one is a robust spoken language analysis part, and the other is a rule based machine translation (RBMT) part. The RBMT technology has been successful in commercial business with its ease of maintenance of translation quality and coverage. In contract, it rest on reliable analysis methods. So we proposed a robust parsing method for spoken language [5]. In our proposal, we treat input- utterances as a sequence of fragmental phrases to achieve high acceptability. Then our robust grammar and parser unifies fragmental phrases in several candidates of coherent syntactic units, and enumerates all possible combinations as interpretations. Each interpretation is tested according to syntactic/semantic preferences and the optimum interpretation is to be translated. But it is difficult to determine whether each fragment should be treated separately or not. This sort of problem is common to all spoken language applications. One key technique is that how to associate each fragment with a suitable unit for a practical application. We regard a clause and dependency structure as one of a clue, since a clause and its dependency represents a syntactic/semantic unit even for a spoken language. In this paper, we propose new analysis method that estimates clause structures and evaluates validities combined with dependency preferences. Our method is designed to work on such a syntax forest like structure. It allows to evaluate all candidates efficiently and to choose totally optimum one. 2 The forest driven spoken language translation 2.1 Overview of our translation process Figure 1 shows the outline of our SLT system based on the “forest driven spoken language analysis” method [1, 5]. It works based on a packed shared forest structure (syntax forest) generated by extended 265 2 Satoshi Kamatani, Kentaro …

Content truncated to 500 words