Collocational Translation Memory Extraction Based on Statistical and Linguistic Information

Collocational Translation Memory Extraction Based on Statistical and Linguistic Information Jia-Yan Jian Department of Computer Science National Tsing Hua University 101, Kuangfu Road, Hsinchu, Taiwan d914339@oz.nthu.edu.tw Yu-Chia Chang Inst. of Information System and Applictaion National Tsing Hua University 101, Kuangfu Road, Hsinchu, Taiwan u881222@alumni.nthu.edu.tw Jason S. Chang Department of Computer Science National Tsing Hua University 101, Kuangfu Road, Hsinchu, Taiwan jschang@cs.nthu.edu.tw Abstract. In this paper, we propose a new method for bilingual collocation extraction from a parallel corpus to provide phrasal translation memory. The method integrates statistical and linguistic information for effective extraction of collocations. The linguistic information includes parts of speech, chunks, and clauses. With an implementation of the method, we obtain first an extended list of collocations from monolingual corpora such as British National Corpus (BNC). Subsequently, we exploit the list to identify English collocations in Sinorama Parallel Corpus (SPC). Finally, we use word alignment techniques to retrieve the translation equivalent of English collocations from the bilingual corpus, so as to provide phrasal translation memory for machine translation system. Based on the strength of chunk and clause analyses, we are able to extract a large number of collocations and translations with much less time and effort than those required by N-gram analysis or full parsing. Furthermore, we also consider longer collocation pattern such as a preposition involved in VN collocation. In the future, we plan to extend the method to other types of collocation. Keyword. Bilingual Collocation Extraction, Collocational Translation Memory, Collocational Concordancer 1 Introduction Example-based machine translation (EBMT), a corpus-based MT method, has been recently suggested as an efficient step toward automatic translation (Nagao, 1981; Kitano, 1993, Carl, 1999, Andrimanankasian et al., 1999; Brown, 2000). Under the approach, systems exploited examples similar to input and adjusted the translations to obtain the result. Translations are preprocessed and stored in a translation memory which serves as an archive of existing translation for MT system to reuse. Nowadays, there have been a number of transducers applied to convert sentences in bilingual corpus into translation patterns, which can be further exploited as a translation memory, such as Transit1, Deja–Vu2, TransSearch3, TOTALrecall4, and so on. A problem that most MT system may encounter is the collocational translation if the system intends not to literally translate the input text. This smaller syntax unit not only facilitates a more native-like translation, but also enhances the performance of recent EBMT system. Elastic collocation structure provides more flexibility in handle translation pattern as in “…not yet to take what he wants into consideration…” 1 Transit (http://www.star-group.net/eng/software/sprachtech/transit.html) 2 Deja–Vu (http://www.atril.com/) 3 TransSearch (http://www.tsrali.com/) 4 TOTALrecall(http://candle.cs.nthu.edu.tw/Counter/Counter.asp?funcID=1) 2 Extraction of Collocational Translation Memory Using valuable linguistic information—chunk and clause analyses, we can retrieve Verb-Noun collocations from a large corpus (i.e. BNC) with good quality and quantity. We further use this collocation type list to identify the concise collocational instances in a bilingual corpus (i.e. SPC). We also use word-alignment technique to extract the matching translation of verb and noun respectively, so as to obtain phrasal translation memory. The detailed …

Content truncated to 500 words