Gappy Translation Units under Left-to-Right SMT Decoding

Proceedings of the 13th Annual Conference of the EAMT, pages 66–73, Barcelona, May 2009 Gappy Translation Units under Left-to-Right SMT Decoding Josep M. Crego Spoken Language Processing Group LIMSI-CNRS, BP 133 91430 Orsay cedex, France jmcrego@limsi.fr Fran¸cois Yvon Univ Paris-Sud 11 LIMSI-CNRS, BP 133 91430 Orsay cedex, France yvon@limsi.fr Abstract This paper presents an extension for a bilingual n-gram statistical machine translation (SMT) system based on al- lowing translation units with gaps. Our gappy translation units can be seen as a first step towards introducing hierar- chical units similar to those employed in hierarchical MT systems. Our goal is double. On the one hand we aim at capturing the benefits of the higher generalization power shown by hierar- chical systems. On the other hand, we want to avoid the computational bur- den of decoding based on parsing tech- niques, which among other drawbacks, make difficult the introduction of the required target language model costs. Our experiments show slight but con- sistent improvements for Chinese-to- English machine translation. Accu- racy results are competitive with those achieved by a state-of-the-art phrase- based system. 1 Introduction Work in SMT has evolved from the traditional word-based (Brown et al., 1993) to the cur- rent phrase-based (Och et al., 1999; Zens et al., 2002; Koehn et al., 2003) and hierarchical- based (Melamed, 2004; Chiang, 2007) trans- lation models. Phrase-based and hierarchical systems are also characterized by the underly- ing formal device employed to produce transla- tions (Knight, 2008): finite-state transducers (FST) on the one hand, and tree transducers c⃝ 2009 European Association for Machine Translation. (TT) on the other hand, specified respectively by rational and context-free grammars, thus implying clear differences in generative power. A thorough comparison between phrase- based and hierarchical MT can be read in (Zollmann et al., 2008), concluding that hierarchical models slightly outperform phrase-based models under “sufficiently non- monotonic language pairs”. One of the reasons for the gap in performance seems to be the abil- ity to generalize using non-terminal categories beyond the strictly lexicalized knowledge rep- resented in phrase-based models. An illustrative example is given below. consist of the translation from English toIt French of negative verb phrases, which yields the alignment of don’t X ; ne X pas, where X could be replaced by almost any finite verb. In this example, the English token don’t is trans- lated into the French non-contiguous words ne and pas 1. The right translation can only be achieved under phrase-based systems, if X (say want) has been seen in training next to don’t, yielding the translation unit: don′t want : ne veux pas In contrast, under hierarchical systems, it is possible to obtain the right generalization, decomposing the previous pattern as: X → don′t Y : ne Y pas Y → want : veux 1This example is only used for illustrative purposes. The contracted form don’t is not a real issue as most tokenizers split the form as do not, thus solving the alignment problem. 66 This ability to capture better generalization comes …

Content truncated to 500 words