Accurate Non-Hierarchical Phrase-Based Translation

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 966–974, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Accurate Non-Hierarchical Phrase-Based Translation Michel Galley and Christopher D. Manning Computer Science Department Stanford University Stanford, CA 94305 {mgalley,manning}@cs.stanford.edu Abstract A principal weakness of conventional (i.e., non-hierarchical) phrase-based statistical machine translation is that it can only exploit continuous phrases. In this paper, we extend phrase-based decoding to allow both source and target phrasal discontinuities, which provide better generalization toon unseen data and yield significant improvements a standard phrase-based system (Moses). More interestingly, our discontinuous phrase- based system also outperforms a state-of-the-art hierarchical system (Joshua) by a very significant margin (+1.03 BLEU on average on five Chinese- English NIST test sets), even though both Joshua and our system support discontinuous phrases. Since the key difference between these two systems is that ours is not hierarchical—i.e., our system uses a string-based decoder instead of CKY, and it imposes no hard hierarchical reordering constraints during training and decoding—this paper sets out to challenge the commonly held belief that the tree-based parameterization of systems such as Hiero and Joshua is crucial to their good performance against Moses. 1 Introduction Phrase-based machine translation models (Och and Ney, 2004) advanced the state of the art by extend- ing the basic translation unit from words to phrases. By conditioning translations on more than a sin- gle word, a statistical machine translation (SMT) system benefits from the larger context of a phrase pair to properly handle multi-word units and lo- cal reorderings. Experimentally, it was found that longer phrases yield better MT output (Koehn et al., 2003). However, while it is computationally feasi- ble at training time to extract phrase pairs of nearly unbounded size (Zhang and Vogel, 2005; Callison- Burch et al., 2005), phrase pairs applicable at test time tend to be fairly short. Indeed, data sparsity often forces conventional phrase-based systems to segment test sentences into small phrases, and there- fore to translate dependent words (e.g., the French ne . . . pas) separately instead of jointly. We present a solution to this sparsity problem by going beyond using only continuous phrases, and instead define our translation unit as any subset of words of a sentence, i.e., a discontinuous phrase. We generalize conventional multi-beam string-based decoding (Koehn, 2004) to allow variable-size dis- continuities in both source and target phrases. Since each sentence pair can be more flexibly decomposed into translation units, it is possible to exploit the rich context of longer (possibly discontinuous) phrases to improve translation quality. Our decoder provides two extensions to Moses (Koehn et al., 2007): (a) to cope with source gaps, we follow (Lopez, 2007) to efficiently find all discontinuous phrases in the train- ing data that also appear in the input sentence; (b) to enable target discontinuities, we augment transla- tion hypotheses to not only record the current par- tial translation, but also a set of subphrases that may be appended to the partial …

Content truncated to 500 words