Segment Choice Models: Feature-Rich Models for Global Distortion in Statistical Machine Translation

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 25–32, New York, June 2006. c⃝2006 Association for Computational Linguistics Segment Choice Models: Feature-Rich Models for Global Distortion in Statistical Machine Translation Roland Kuhn, Denis Yuen, Michel Simard, Patrick Paul, George Foster, Eric Joanis, and Howard Johnson Institute for Information Technology, National Research Council of Canada Gatineau, Québec, CANADA Email: {Roland.Kuhn, Michel.Simard, Patrick.Paul, George.Foster, Eric.Joanis, Howard.Johnson}@cnrc-nrc.gc.ca; Denis Yuen: mucous@gmail.com Abstract This paper presents a new approach to distortion (phrase reordering) in phrase- based machine translation (MT). Distor- tion is modeled as a sequence of choices during translation. The approach yields trainable, probabilistic distortion models that are global: they assign a probability to each possible phrase reordering. These “segment choice” models (SCMs) can be trained on “segment-aligned” sentence pairs; they can be applied during decoding or rescoring. The approach yields a metric called “distortion perplexity” (“disperp”) for comparing SCMs offline on test data, analogous to perplexity for language models. A decision-tree-based SCM is tested on Chinese-to-English translation, and outperforms a baseline distortion penalty approach at the 99% confidence level. 1 Introduction: Defining SCMs The work presented here was done in the context of phrase-based MT (Koehn et al., 2003; Och and Ney, 2004). Distortion in phrase-based MT occurs when the order of phrases in the source-language sentence changes during translation, so the order of corresponding phrases in the target-language trans- lation is different. Some MT systems allow arbi- trary reordering of phrases, but impose a distortion penalty proportional to the difference between the new and the original phrase order (Koehn, 2004). Some interesting recent research focuses on reor- dering within a narrow window of phrases (Kumar and Byrne, 2005; Tillmann and Zhang, 2005; Till- mann, 2004). The (Tillmann, 2004) paper intro- duced lexical features for distortion modeling. A recent paper (Collins et al., 2005) shows that major gains can be obtained by constructing a parse tree for the source sentence and then applying hand- crafted reordering rules to rewrite the source in target-language-like word order prior to MT. Our model assumes that the source sentence is completely segmented prior to distortion. This simplifying assumption requires generation of hy- potheses about the segmentation of the complete source sentence during decoding. The model also assumes that each translation hypothesis grows in a predetermined order. E.g., Koehn’s decoder (Koehn 2004) builds each new hypothesis by add- ing phrases to it left-to-right (order is deterministic for the target hypothesis). Our model doesn’t re- quire this order of operation – it would support right-to-left or inwards-outwards hypothesis con- struction – but it does require a predictable order. One can keep track of how segments in the source sentence have been rearranged during de- coding for a given hypothesis, using what we call a “distorted source-language hypothesis” (DSH). A similar concept appears in (Collins et al., 2005) (this paper’s preoccupations strongly resemble 25 ours, though our method is completely different: we don’t parse the source, and use only automati- cally generated …

Content truncated to 500 words