Improved Models of Distortion Cost for Statistical Machine Translation

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 867–875, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Improved Models of Distortion Cost for Statistical Machine Translation Spence Green, Michel Galley, and Christopher D. Manning Computer Science Department Stanford University Stanford, CA 94305 {spenceg,mgalley,manning}@stanford.edu Abstract The distortion cost function used in Moses- style machine translation systems has two flaws. First, it does not estimate the future cost of known required moves, thus increas- ing search errors. Second, all distortion is penalized linearly, even when appropriate re- orderings are performed. Because the cost function does not effectively constrain search, translation quality decreases at higher dis- tortion limits, which are often needed when translating between languages of different ty- pologies such as Arabic and English. To ad- dress these problems, we introduce a method for estimating future linear distortion cost, and a new discriminative distortion model that pre- dicts word movement during translation. In combination, these extensions give a statis- tically significant improvement over a base- line distortion parameterization. When we triple the distortion limit, our model achieves a +2.32 BLEU average gain over Moses. 1 Introduction It is well-known that translation performance in Moses-style (Koehn et al., 2007) machine transla- tion (MT) systems deteriorates when high distortion is allowed. The linear distortion cost model used in these systems is partly at fault. It includes no es- timate of future distortion cost, thereby increasing the risk of search errors. Linear distortion also pe- nalizes all re-orderings equally, even when appro- priate re-orderings are performed. Because linear distortion, which is a soft constraint, does not effec- tively constrain search, a distortion limit is imposed on the translation model. But hard constraints are ultimately undesirable since they prune the search space. For languages with very different word or- Followers of all of the Christian and Islamic sects engaged ���� Verb �� NP-OBJ ���������� PP ���������������������������������� NP-SBJ for them in waiting Figure 1: The oracle translation for this Arabic VOS sen- tence would be pruned during search using typical dis- tortion parameters. The Arabic phrases read right-to-left, but we have ordered the sentence from left-to-right in or- der to clearly illustrate the re-ordering problem. ders in which significant re-ordering is required, the distortion limit can eliminate the oracle, or “best,” translation prior to search, placing an artificial limit on translation performance (Auli et al., 2009). To illustrate this problem, consider the Arabic- English example in Figure 1. Assuming that the En- glish translation is constructed left-to-right, the verb ���� shaaraka must be translated after the noun phrase (NP) subject. If P phrases are used to trans- late the Arabic source s to the English target t, then the (unsigned) linear distortion is given by D(s, t) = p1 first + P � i=2 ��pi−1 last + 1 − pi first �� (1) where pfirst and plast are the first and last source word indices, respectively, in phrase i. By this for- mula, the cost of the …

Content truncated to 500 words