Phrase Reordering Model Integrating Syntactic Knowledge for SMT

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 533–540, Prague, June 2007. c⃝2007 Association for Computational Linguistics Phrase Reordering Model Integrating Syntactic Knowledge for SMT Dongdong Zhang, Mu Li, Chi-Ho Li, Ming Zhou Microsoft Research Asia Beijing, China {dozhang,muli,chl,mingzhou}@microsoft.com Abstract Reordering model is important for the sta- tistical machine translation (SMT). Current phrase-based SMT technologies are good at capturing local reordering but not global reordering. This paper introduces syntactic knowledge to improve global reordering capability of SMT system. Syntactic know- ledge such as boundary words, POS infor- mation and dependencies is used to guide phrase reordering. Not only constraints in syntax tree are proposed to avoid the reor- dering errors, but also the modification of syntax tree is made to strengthen the capa- bility of capturing phrase reordering. Fur- thermore, the combination of parse trees can compensate for the reordering errors caused by single parse tree. Finally, expe- rimental results show that the performance of our system is superior to that of the state-of-the-art phrase-based SMT system. 1 Introduction In the last decade, statistical machine translation (SMT) has been widely studied and achieved good translation results. Two kinds of SMT system have been developed, one is phrase-based SMT and the other is syntax-based SMT. In phrase-based SMT systems (Koehn et al., 2003; Koehn, 2004), foreign sentences are firstly segmented into phrases which consists of adjacent words. Then source phrases are translated into tar- get phrases respectively according to knowledge usually learned from bilingual parallel corpus. Fi- nally the most likely target sentence based on a certain statistical model is inferred by combining and reordering the target phrases with the aid of search algorithm. On the other hand, syntax-based SMT systems (Liu et al., 2006; Yamada et al., 2001) mainly depend on parse trees to complete the translation of source sentence. Figure 1: A reordering example As studied in previous SMT projects, language model, translation model and reordering model are the three major components in current SMT sys- tems. Due to the difference between the source and target languages, the order of target phrases in the target sentence may differ from the order of source phrases in the source sentence. To make the trans- lation results be closer to the target language style, a mathematic model based on the statistic theory is constructed to reorder the target phrases. This sta- tistic model is called as reordering model. As shown in Figure 1, the order of the translations of “欧元” and “的” is changed. The order of the IP VP ADVP NP DNP VP NP NN AD DEG 欧元 VV 的 大幅 升值 the Euro the significant appreciation of 533 translation of “欧元/的” and “大幅/升值” is al- tered as well. The former reordering case with the smaller distance is usually referred as local reor- dering and the latter with the longer distance reor- dering as global reordering. Phrase-based SMT system can effectively capture the local word reor- dering information which is …

Content truncated to 500 words