Automatic Translation Template Acquisition Based on Bilingual Structure Alignment

83 Computational Linguistics and Chinese Language Processing Vol. 6, No. 1, February 2001, pp. 83-108  Computational Linguistics Society of R.O.C. Automatic Translation Template Acquisition Based on Bilingual Structure Alignment1 Yajuan Lü *, Ming Zhou +, Sheng Li *, Changning Huang +, Tiejun Zhao * Abstract Knowledge acquisition is a bottleneck in machine translation and many NLP tasks. A method for automatically acquiring translation templates from bilingual corpora is proposed in this paper. Bilingual sentence pairs are first aligned in syntactic structure by combining a language parsing with a statistical bilingual language model. The alignment results are used to extract translation templates which turn out to be very useful in real machine translation. Keywords: Bilingual corpus, Translation template acquisition, Structure alignment, Machine translation 1. Introduction Bilingual corpora have been recognized as a valuable resource for knowledge acquisition in machine translation and many other NLP tasks. To make better use of them, bilingual corpora are often aligned first. Intensive researches have been done on sentence and word level alignment [Brown et al. 1991, Church 1993, Ker et al. 1997, Huang et al. 2000]. These alignments have been proven to be very useful in machine translation, word sense disambiguation, information retrieval, translation lexicon extraction, and so on. With a sentence aligned parallel English-Chinese corpus ready in hand, this paper extends word-level alignment to syntactic structure alignment with the aim of acquiring structural translation templates automatically. 1 The work was partially done at Microsoft Research Asia. * Dept. of Computer Science & Engineering, Harbin Institute of Technology, Harbin, 15001 E-mail: Lü Yajuan: lyj@mtlab.hit.edu.cn Zhao Tiejun: tjzhao@mtlab.hit.edu.cn + Microsoft Research Asia, Beijing, 10080 E-mail: Zhou Ming: mingzhou@microsoft.com Huang Changning: cnhuang@microsoft.com 84 Y. Lü et al. Numerous researches have been done to acquire knowledge from bilingual corpora. Many of these studies aimed to acquire word or phrase translation lexicons [Shin et al. 1996, Fung et al. 1997, Ralf 1997, Turcato 1998]. This paper focuses on the automatic learning of translation templates, which are especially useful for machine translation. In [Guvenir et al. 1998], [Malavazos et al. 2000] and [Cicekli et al. 2001], analogical models were proposed to learn translation templates. By grouping similar translation examples and replacing their difference with a variable, they could obtain translation templates. Structure alignment has been studied by several researchers for use in structural translation template acquisition. Most of the approaches have followed what may be called a “parse-parse-match” procedure [Wu 1997]. The main idea is that each language of the parallel corpus is first parsed individually using a monolingual grammar, and then the corresponding constituents are matched using some heuristic procedures. The works by [Kaji et al. 1992], [Almuallim et al. 1994], [Grishman et al. 1994], [Matsumoto et al. 1995], [Meyers et al. 1998], [Watanabe et al. 2000] etc. can be considered such approaches. Differences between them are in their parsing grammars and heuristic procedures. Kaji and Watanabe used phrase structure grammar, while Grishman employed a regularized syntactic structure. The dependency structure is used in most of the other …

Content truncated to 500 words