A Probabilistic Approach to Syntax-based Reordering for Statistical Machine Translation

Proceedings of the 45th Annual Meeting of the Association of Computational Linguistics, pages 720–727, Prague, Czech Republic, June 2007. c⃝2007 Association for Computational Linguistics A Probabilistic Approach to Syntax-based Reordering for Statistical Machine Translation Chi-Ho Li, Dongdong Zhang, Mu Li, Ming Zhou Microsoft Research Asia Beijing, China chl, dozhang@microsoft.com muli, mingzhou@microsoft.com Minghui Li, Yi Guan Harbin Institute of Technology Harbin, China mhli@insun.hit.edu.cn guanyi@insun.hit.edu.cn Abstract Inspired by previous preprocessing ap- proaches to SMT, this paper proposes a novel, probabilistic approach to reordering which combines the merits of syntax and phrase-based SMT. Given a source sentence and its parse tree, our method generates, by tree operations, an n-best list of re- ordered inputs, which are then fed to stan- dard phrase-based decoder to produce the optimal translation. Experiments show that, for the NIST MT-05 task of Chinese-to- English translation, the proposal leads to BLEU improvement of 1.56%. 1 Introduction The phrase-based approach has been considered the default strategy to Statistical Machine Translation (SMT) in recent years. It is widely known that the phrase-based approach is powerful in local lexical choice and word reordering within short distance. However, long-distance reordering is problematic in phrase-based SMT. For example, the distance- based reordering model (Koehn et al., 2003) al- lows a decoder to translate in non-monotonous or- der, under the constraint that the distance between two phrases translated consecutively does not ex- ceed a limit known as distortion limit. In theory the distortion limit can be assigned a very large value so that all possible reorderings are allowed, yet in practise it is observed that too high a distortion limit not only harms efficiency but also translation per- formance (Koehn et al., 2005). In our own exper- iment setting, the best distortion limit for Chinese- English translation is 4. However, some ideal trans- lations exhibit reorderings longer than such distor- tion limit. Consider the sentence pair in NIST MT- 2005 test set shown in figure 1(a): after translating the word “��/mend”, the decoder should ‘jump’ across six words and translate the last phrase “� � ��/fissures in the relationship”. Therefore, while short-distance reordering is under the scope of the distance-based model, long-distance reorder- ing is simply out of the question. A terminological remark: In the rest of the paper, we will use the terms global reordering and local reordering in place of long-distance reordering and short-distance reordering respectively. The distinc- tion between long and short distance reordering is solely defined by distortion limit. Syntax1 is certainly a potential solution to global reordering. For example, for the last two Chinese phrases in figure 1(a), simply swapping the two chil- dren of the NP node will produce the correct word order on the English side. However, there are also reorderings which do not agree with syntactic anal- ysis. Figure 1(b) shows how our phrase-based de- coder2 obtains a good English translation by reorder- ing two blocks. It should be noted that the second Chinese block “�� �” and its English counterpart “at the end of” are not …

Content truncated to 500 words