Recognizing Coordinate Structures for Machine Translation of English Patent Documents

Recognizing Coordinate Structures for Machine Translation of English Patent Documents* Yoon-Hyung Roh, Ki-Young Lee, Sung-Kwon Choi, Oh-Woog Kwon, and Young-Gil Kim Natural Language Processing Research Team, Electronics and Telecommunications Research Institute, 161 Gajeong-dong, Yuseong-gu, Daejeon, 305-350, Korea {yhroh, leeky, choisk, ohwoog, kimyk}@etri.re.kr Abstract. Patent machine translation is one of main target areas of current practical MT systems. Patent documents have their own peculiar description style. Especially, abstracts or claims in patent documents are characterized by their long and complex syntactic structures, which are often caused by coordination. So, syntactic analysis of patent documents requires special treatment for coordination. This paper describes a method to deal with long sentences in patent documents by recognizing coordinate structures. Coordinate structures are recognized using a similarity table which reflects parallelism between conjuncts. Our method is applied to a practical MT system and improves its quality and efficiency. Keywords: Machine Translation, Patent Document, Coordinate Structure, Syntactic Analysis 1. Introduction Patent machine translation is one of main target areas of current practical MT systems such as the English-Korean patent machine translation system (Kwon, 2007). Patent documents have their own description style and there have been some studies on the analysis of patent documents (Sheremetyeva, 2003; Shinmori and Okumura, 2003; Shinmori and Okumura, 2002). Especially, abstracts or claims in the patent documents are notorious for their long and complex syntactic structures, which are usually caused by coordination or relative clauses. Long sentences formed by relative clauses can be handled by segmentation (Kim et al, 2001). On the other hand, in case of long sentences formed by coordination, segmentation can cause syntactic analysis errors, because a segment resulting from segmentation can be dependent on the other constituents in the parse tree. Also, coordinate structures in patent documents have usually a large number of coordinate conjuncts (which we will call nodes) and can cause syntactic ambiguity explosion and parsing failure at the worst case in practical MT systems. So, syntactic analysis of patent documents requires special treatment for coordination. There have been many computational researches about coordinate structures (Kaplan and Maxwell, 1988; Kosy, 1986). But, it is unrealistic to apply most of them to large-scale MT systems as mentioned in (Okumura and Muraki, 1994; Kurohashi and Nagao,. 1994). The more practical approaches about analyzing coordinate structures such as (Okumura and Muraki, 1994; * This work was funded by the Ministry of Information and Communication of Korean government. Copyright 2008 by Yoon-Hyung Roh, Ki-Young Lee, Sung-Kwon Choi, Oh-Woog Kwon, and Young- Gil Kim 460 22nd Pacific Asia Conference on Language, Information and Computation, pages 460–466 Kurohashi and Nagao, 1994; Agarwal and Boggess, 1992) all analyze coordinate structures using parallelism between conjuncts, but they are mainly targeted to recognizing two conjuncts (i.e., pre-conjuct and post-conjuct). Out of them, (Kurohashi and Nagao,. 1994) is one of the most practical approaches. They recognize coordinate structures in a Japanese sentence by constructing a similarity matrix between bunsetsus and searching a path with the highest parallelism in the similarity matrix using a dynamic programming method. …

Content truncated to 500 words