Graph-based Learning for Statistical Machine Translation

Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the ACL, pages 119–127, Boulder, Colorado, June 2009. c⃝2009 Association for Computational Linguistics Graph-based Learning for Statistical Machine Translation Andrei Alexandrescu Dept. of Comp. Sci. Eng. University of Washington Seattle, WA 98195, USA andrei@cs.washington.edu Katrin Kirchhoff Dept. of Electrical Eng. University of Washington Seattle, WA 98195, USA katrin@ee.washington.edu Abstract Current phrase-based statistical machine translation systems process each test sentence in isolation and do not enforce global consis- tency constraints, even though the test data is often internally consistent with respect to topic or style. We propose a new consistency model for machine translation in the form of a graph-based semi-supervised learning algorithm that exploits similarities between training and test data and also similarities between different test sentences. The algo- rithm learns a regression function jointly over training and test data and uses the resulting scores to rerank translation hypotheses. Eval- uation on two travel expression translation tasks demonstrates improvements of up to 2.6 BLEU points absolute and 2.8% in PER. 1 Introduction Current phrase-based statistical machine translation (SMT) systems commonly operate at the sentence level—each sentence is translated in isolation, even when the test data consists of internally coherent paragraphs or stories, such as news articles. For each sentence, SMT systems choose the translation hypothesis that maximizes a combined log-linear model score, which is computed independently of all other sentences, using globally optimized com- bination weights. Thus, similar input strings may be translated in very different ways, depending on which component model happens to dominate the combined score for that sentence. This is illustrated by the following example (from the IWSLT 2007 Arabic-English translation task): Source 1: Asf lA ymknk *lk hnAk klfp HwAly vmAnyn dwlAr lAlsAEp AlwAHdp Ref: sorry you can’t there is a cost the charge is eighty dollars per hour 1-best: i’m sorry you can’t there in the cost about eighty dollars for a one o’clock Source 2: E*rA lA ymknk t$gyl AltlfAz HtY tqlE AlTA}rp Ref: sorry you cannot turn the tv on until the plane has taken off 1-best: excuse me i you turn tv until the plane departs The phrase lA ymknk (you may not/you cannot) is translated differently (and wrongly in the sec- ond case) due to different segmentations and phrase translations chosen by the decoder. Though differ- ent choices may be sometimes appropriate, the lack of constraints enforcing translation consistency of- ten leads to suboptimal translation performance. It would be desirable to counter this effect by encour- aging similar outputs for similar inputs (under a suit- ably defined notion of similarity, which may include e.g. a context specification for the phrase/sentence). In machine learning, the idea of forcing the out- puts of a statistical learner to vary smoothly with the underlying structure of the inputs has been formal- ized in the graph-based learning (GBL) framework. In GBL, both labeled (train) and unlabeled (test) data samples are jointly represented as vertices in a graph whose edges encode pairwise similarities be- …

Content truncated to 500 words