Using a Dependency Parser to Improve SMT for Subject-Object-Verb Languages

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 245–253, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Using a Dependency Parser to Improve SMT for Subject-Object-Verb Languages Peng Xu, Jaeho Kang, Michael Ringgaard and Franz Och Google Inc. 1600 Amphitheatre Parkway Mountain View, CA 94043, USA {xp,jhkang,ringgaard,och}@google.com Abstract We introduce a novel precedence reordering approach based on a dependency parser to sta- tistical machine translation systems. Similar to other preprocessing reordering approaches, our method can efficiently incorporate linguis- tic knowledge into SMT systems without in- creasing the complexity of decoding. For a set of five subject-object-verb (SOV) order lan- guages, we show significant improvements in BLEU scores when translating from English, compared to other reordering approaches, in state-of-the-art phrase-based SMT systems. 1 Introduction Over the past ten years, statistical machine transla- tion has seen many exciting developments. Phrase- based systems (Och, 2002; Koehn et.al., 2003; Och and Ney, 2004) advanced the machine transla- tion field by allowing translations of word sequences (a.k.a., phrases) instead of single words. This ap- proach has since been the state-of-the-art because of its robustness in modeling local word reordering and the existence of an efficient dynamic programming decoding algorithm. However, when phrase-based systems are used between languages with very different word or- ders, such as between subject-verb-object (SVO) and subject-object-verb (SOV) languages, long dis- tance reordering becomes one of the key weak- nesses. Many reordering methods have been pro- posed in recent years to address this problem in dif- ferent aspects. The first class of approaches tries to explicitly model phrase reordering distances. Distance based distortion model (Och, 2002; Koehn et.al., 2003) is a simple way of modeling phrase level reordering. It penalizes non-monotonicity by applying a weight to the number of words between two source phrases corresponding to two consecutive target phrases. Later on, this model was extended to lexicalized phrase reordering (Tillmann, 2004; Koehn, et.al., 2005; Al-Onaizan and Papineni, 2006) by applying different weights to different phrases. Most recently, a hierarchical phrase reordering model (Galley and Manning, 2008) was proposed to dynamically deter- mine phrase boundaries using efficient shift-reduce parsing. Along this line of research, discrimina- tive reordering models based on a maximum entropy classifier (Zens and Ney, 2006; Xiong, et.al., 2006) also showed improvements over the distance based distortion model. None of these reordering models changes the word alignment step in SMT systems, therefore, they can not recover from the word align- ment errors. These models are also limited by a maximum allowed reordering distance often used in decoding. The second class of approaches puts syntactic analysis of the target language into both modeling and decoding. It has been shown that direct model- ing of target language constituents movement in ei- ther constituency trees (Yamada and Knight, 2001; Galley et.al., 2006; Zollmann et.al., 2008) or depen- dency trees (Quirk, et.al., 2005) can result in signifi- cant improvements in translation quality for translat- ing languages like Chinese and Arabic into English. A simpler alternative, …

Content truncated to 500 words