Automatically Learning Source-side Reordering Rules for Large Scale Machine Translation

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 376–384, Beijing, August 2010 Automatically Learning Source-side Reordering Rules for Large Scale Machine Translation Dmitriy Genzel Google, Inc. dmitriy@google.com Abstract We describe an approach to automatically learn reordering rules to be applied as a preprocessing step in phrase-based ma- chine translation. We learn rules for 8 dif- ferent language pairs, showing BLEU im- provements for all of them, and demon- strate that many important order trans- formations (SVO to SOV or VSO, head- modifier, verb movement) can be captured by this approach. 1 Introduction One of the major problems of modern statisti- cal machine translation relates to its difficulties in producing the correct word order on the target side of the translation where the source side or- der is not the same as the target side. In many cases where the translation is spectacularly bad, if one only enters the source sentence in the word or- der of the target language the translation becomes near-perfect (largely because the language model can now make sense of it). The word order prob- lems are especially extensive for languages that have major differences, such as SOV vs. SVO languages, but also cause insidious, but entirely avoidable errors for the language pairs where the word order is almost right, but not quite1. For practical reasons all phrase-based decoders limit the amount of reordering allowed and thus are completely unable to produce correct translations when the necessary movement is over a large dis- tance. Furthermore, where the actual systematic reordering for the two languages is within the de- coder’s search space, it is penalized just as any 1For example of the latter kind, verb movement for English-German and similar language pairs often causes verbs to be aligned to nothing and to be altogether dropped in translation. other kind of reordering, whereas doing anything other than this systematic reordering should in fact be penalized. It has been argued that this is a fundamental flaw in phrase-based decoding systems and hier- archical and syntax-based systems have been pro- posed to solve this problem. These systems can in principle resolve a part of this problem, but at a significant time cost during training, and even worse, during translation, making it less practical for realtime systems. Instead we propose a system for learning pre-ordering rules automatically from data and demonstrate that it can capture many dif- ferent kinds of reordering phenomena and do so at no additional online cost. 2 Related Work Many solutions to the reordering problem have been proposed, e.g. syntax-based models (Chi- ang, 2005), lexicalized reordering (Och et al., 2004), and tree-to-string methods (Zhang et al., 2006). All these methods try to solve the reorder- ing problem in different ways, but have the fol- lowing problems in common: word alignment is not affected by them and they tend to introduce significant additional work to be done at transla- tion time. Most state of the art systems use HMM or IBM Model 4 word alignment, …

Content truncated to 500 words