Extraction of Broad-Scale, High-Precision Japanese-English Parallel Translation Expressions Using Lexical Information and Rules

Extraction of Broad-Scale, High-Precision Japanese-English Parallel Translation Expressions Using Lexical Information and Rules ⋆ Qing Ma a, Shinya Sakagami a, and Masaki Murata b aDepartment of Applied Mathematics and Informatics, Ryukoku University, 1-5 Oei-Yokoya, Seta, Otsu 520-2194, Japan qma@math.ryukoku.ac.jp bDepartment of Information and Electronics, Tottori University, 4-101 Koyama-Minami, Tottori 680-8552, Japan murata@ike.tottori-u.ac.jp Abstract. Extraction was attempted of broad-scale, high-precision Japanese-English paral- lel translation expressions from large aligned parallel corpora. To acquire broad-scale parallel translation expressions, a new method was used to extract single Japanese and English word n-grams, by which as many parallel translation expressions as possible could then be ex- tracted. To achieve high extraction precision, first, hand-crafted rules were used to prune the unnecessary words often found in expressions extracted on the basis of word n-grams, and lexical information was used to refine the parallel translation expressions. Computer exper- iments with aligned parallel corpora consisting of about 280,000 pairs of Japanese-English parallel sentences found that more than 125,000 pairs of parallel translation expressions could be extracted with a precision of 0.96. These figures show that the proposed methods for ex- tracting a broad range of parallel translation expressions have reached a level high enough for practical use. Keywords: parallel translation expression, extraction, rule, lexical information, English- writing support 1 Introduction Non-native speakers often have problems explaining ideas or presenting achievements in writ- ten English, in part because of the large amount of time needed to determine which possible translation of an expression most suits the context. Trying to develop English-writing support tools that will enable non-native speakers to produce nearly perfect English sentences for mixed English-Japanese sentences–in which expressions without know translations are simply written in Japanese–Ma et al. (2008; 2009) have developed systems that can provide support at the word and phrase levels. That is, the given Japanese parts in the mixed English-Japanese sentences can be words or phrases. For phrase-level support in those systems, the Japanese parts are extracted from the mixed sentences and segmented into words, the candidate English equivalents of the segmented Japanese words are identified by searching through a Japanese-English dictionary, and the best equivalents of the Japanese phrases are selected from the combinations of the candidate translations of the single words. This kind of support, however, has two problems. One is the variety of the phrase patterns that the system can support is limited because the English phrases are generated only from the combinations of the candidate translations of single words. The other is the processing is time-consuming because a large number of the translations of single words results in an enormous number of combinations. ⋆ This work was supported by a Grant-in-Aid for Scientific Research (KAKENHI (C) 19500133) from the Ministry of Education, Culture, Sports, Science and Technology, Japan. Copyright 2011 by Qing Ma, Shinya Sakagami, and Masaki Murata 25th Pacific Asia Conference on Language, Information and Computation, pages 577–586 577 We think that an effective way to solve these problems would be to use an approach based on translation patterns: …

Content truncated to 500 words