Learning Bilingual Linguistic Reordering Model for Statistical Machine Translation

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 254–262, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Learning Bilingual Linguistic Reordering Model for Statistical Machine Translation Han-Bin Chen, Jian-Cheng Wu and Jason S. Chang Department of Computer Science National Tsing Hua University 101, Guangfu Road, Hsinchu, Taiwan {hanbin,d928322,jschang}@cs.nthu.edu.tw Abstract In this paper, we propose a method for learn- ing reordering model for BTG-based statisti- cal machine translation (SMT). The model focuses on linguistic features from bilingual phrases. Our method involves extracting reor- dering examples as well as features such as part-of-speech and word class from aligned parallel sentences. The features are classified with special considerations of phrase lengths. We then use these features to train the maxi- mum entropy (ME) reordering model. With the model, we performed Chinese-to-English translation tasks. Experimental results show that our bilingual linguistic model outper- forms the state-of-the-art phrase-based and BTG-based SMT systems by improvements of 2.41 and 1.31 BLEU points respectively. 1 Introduction Bracketing Transduction Grammar (BTG) is a spe- cial case of Synchronous Context Free Grammar (SCFG), with binary branching rules that are either straight or inverted. BTG is widely adopted in SMT systems, because of its good trade-off be- tween efficiency and expressiveness (Wu, 1996). In BTG, the ratio of legal alignments and all possi- ble alignment in a translation pair drops drastically especially for long sentences, yet it still covers most of the syntactic diversities between two lan- guages. It is common to utilize phrase translation in BTG systems. For example in (Xiong et al., 2006), source sentences are segmented into phrases. Each sequences of consecutive phrases, mapping to cells in a CKY matrix, are then translated through a bi- lingual phrase table and scored as implemented in (Koehn et al., 2005; Chiang, 2005). In other words, their system shares the same phrase table with standard phrase-based SMT systems. 3 年 前 3 年 後 three after A2 years 1A three ago A2 years 1A (a) (b) Figure 1: Two reordering examples, with straight rule applied in (a), and inverted rule in (b). On the other hand, there are various proposed BTG reordering models to predict correct orienta- tions between neighboring blocks (bilingual phrases). In Figure 1, for example, the role of reor- dering model is to predict correct orientations of neighboring blocks A1 and A2. In flat model (Wu, 1996; Zens et al., 2004; Kumar and Byrne, 2005), reordering probabilities are assigned uniformly during decoding, and can be tuned depending on different language pairs. It is clear, however, that this kind of model would suffer when the dominant rule is wrongly applied. Predicting orientations in BTG depending on context information can be achieved with lexical features. For example, Xiong et al. (2006) pro- posed MEBTG, based on maximum entropy (ME) classification with words as features. In MEBTG, first words of blocks are considered as the fea- tures, which are then used to train a ME model 254 for predicting orientations of neighboring …

Content truncated to 500 words