Empirical Study of Utilizing Morph-Syntactic Information in SMT

Empirical Study of Utilizing Morph-Syntactic Information in SMT Young-Sook Hwang, Taro Watanabe, and Yutaka Sasaki ATR SLT Research Labs, 2-2-2 Hikaridai Seika-cho, Soraku-gun Kyoto, 619-0288, Japan {youngsook.hwang, taro.watanabe, yutaka.sasaki}@atr.jp Abstract. In this paper, we present an empirical study that utilizes morph-syntactical information to improve translation quality. With three kinds of language pairs matched according to morph-syntactical similar- ity or difference, we investigate the effects of various morpho-syntactical information, such as base form, part-of-speech, and the relative positional information of a word in a statistical machine translation framework. We learn not only translation models but also word-based/class-based language models by manipulating morphological and relative positional information. And we integrate the models into a log-linear model. Ex- periments on multilingual translations showed that such morphological information as part-of-speech and base form are effective for improving performance in morphologically rich language pairs and that the relative positional features in a word group are useful for reordering the local word orders. Moreover, the use of a class-based n-gram language model improves performance by alleviating the data sparseness problem in a word-based language model. 1 Introduction For decades, many research efforts have contributed to the advance of statisti- cal machine translation. Such an approach to machine translation has proven successful in various comparative evaluations. Recently, various works have im- proved the quality of statistical machine translation systems by using phrase translation [1,2,3,4] or using morpho-syntactic information [6,8]. But most sta- tistical machine translation systems still consider surface forms and rarely use linguistic knowledge about the structure of the languages involved[8]. In this paper, we address the question of the effectiveness of morpho-syntactic features such as parts-of-speech, base forms, and relative positions in a chunk or an ag- glutinated word for improving the quality of statistical machine translations. Basically, we take a statistical machine translation model based on an IBM model that consists of a language model and a separate translation model [5]: eI 1 = argmaxeI 1Pr(f J 1 |eI 1)Pr(eI 1) (1) The translation model links the source language sentence to the target language sentence. The target language model describes the well-formedness of the target language sentence. R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 474–485, 2005. c ⃝ Springer-Verlag Berlin Heidelberg 2005 Empirical Study of Utilizing Morph-Syntactic Information in SMT 475 One of the main problems in statistical machine translation is to learn the less ambiguous correspondences between the words in the source and target languages from the bilingual training data. When translating one source lan- guage(which may be inflectional or non-inflectional) into the morphologically rich language such like Japanese or Korean, the bilingual training data can be exploited better by explicitly taking into account the interdependencies of re- lated inflected or agglutinated forms. In this study, we represent a word with its morphological features in both sides of the source and the target language to learn less ambiguous correspondences between the source and the target lan- guage words or phrases. In addition, we utilize the relative positional information of a …

Content truncated to 500 words