A Simple and Effective Hierarchical Phrase Reordering Model

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 848–856, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics A Simple and Effective Hierarchical Phrase Reordering Model Michel Galley Computer Science Department Stanford University Stanford, CA 94305-9020 galley@cs.stanford.edu Christopher D. Manning Computer Science Department Stanford University Stanford, CA 94305-9010 manning@cs.stanford.edu Abstract While phrase-based statistical machine trans- lation systems currently deliver state-of-the- art performance, they remain weak on word order changes. Current phrase reordering models can properly handle swaps between adjacent phrases, but they typically lack the ability to perform the kind of long-distance re- orderings possible with syntax-based systems. In this paper, we present a novel hierarchical phrase reordering model aimed at improving non-local reorderings, which seamlessly in- tegrates with a standard phrase-based system with little loss of computational efficiency. We show that this model can successfully han- dle the key examples often used to motivate syntax-based systems, such as the rotation of a prepositional phrase around a noun phrase. We contrast our model with reordering models commonly used in phrase-based systems, and show that our approach provides statistically significant BLEU point gains for two language pairs: Chinese-English (+0.53 on MT05 and +0.71 on MT08) and Arabic-English (+0.55 on MT05). 1 Introduction Statistical phrase-based systems (Och and Ney, 2004; Koehn et al., 2003) have consistently de- livered state-of-the-art performance in recent ma- chine translation evaluations, yet these systems re- main weak at handling word order changes. The re- ordering models used in the original phrase-based systems penalize phrase displacements proportion- ally to the amount of nonmonotonicity, with no con- sideration of the fact that some words are far more �� �� �� �� � �� �� �� � ��� �� �� � �� ������������ ���������� �������������� ������������ � � �� � �� � �� � ���������������� ������������� �������������� � �� �� �� � ��� ��� Figure 1: Phase orientations (monotone, swap, discontin- uous) for Chinese-to-English translation. While previous work reasonably models phrase reordering in simple ex- amples (a), it fails to capture more complex reorderings, such as the swapping of “of the region” (b). likely to be displaced than others (e.g., in English-to- Japanese translation, a verb should typically move to the end of the clause). Recent efforts (Tillman, 2004; Och et al., 2004; Koehn et al., 2007) have directly addressed this issue by introducing lexicalized reordering models into phrase-based systems, which condition reordering probabilities on the words of each phrase pair. These models distinguish three orientations with respect to the previous phrase—monotone (M), swap (S), and discontinuous (D)—and as such are primarily de- signed to handle local re-orderings of neighboring phrases. Fig. 1(a) is an example where such a model effectively swaps the prepositional phrase in Luxem- bourg with a verb phrase, and where the noun min- isters remains in monotone order with respect to the previous phrase EU environment. While these lexicalized re-ordering models have shown substantial improvements over unlexicalized phrase-based systems, these models only have a 848 limited ability to capture sensible long distance …

Content truncated to 500 words