Translation model based weighting for phrase extraction

Translation Model Based Weighting for Phrase Extraction Saab Mansour and Hermann Ney Human Language Technology and Pattern Recognition Computer Science Department RWTH Aachen University Aachen, Germany {mansour,ney}@cs.rwth-aachen.de Abstract Domain adaptation for statistical machine translation is the task of altering general models to improve performance on the test domain. In this work, we suggest several novel weighting schemes based on trans- lation models for adapted phrase extrac- tion. To calculate the weights, we first phrase align the general bilingual training data, then, using domain specific transla- tion models, the aligned data is scored and weights are defined over these scores. Ex- periments are performed on two translation tasks, German-to-English and Arabic-to- English translation with lectures as the tar- get domain. Different weighting schemes based on translation models are compared, and significant improvements over auto- matic translation quality are reported. In addition, we compare our work to previ- ous methods for adaptation and show sig- nificant gains. 1 Introduction In recent years, large amounts of monolingual and bilingual training corpora were collected for sta- tistical machine translation (SMT). Early years focused on structured data translation such as newswire and parliamentary discussions. Nowa- days, new domains of translation are being ex- plored, such as talk translation in the IWSLT TED evaluation (Cettolo et al., 2012) and patents trans- lation at the NTCIR PatentMT task (Goto et al., 2013). c⃝ 2014 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. The task of domain adaptation tackles the prob- lem of utilizing existing resources mainly drawn from one domain (e.g. newswire, parliamentary discussion) to maximize the performance on the test domain (e.g. lectures, web forums). The main component of an SMT system is the phrase table, providing the building blocks (i.e. phrase translation pairs) and corresponding trans- lation model scores (e.g., phrase models, word lex- ical smoothing, etc.) to search for the best trans- lation. In this work, we experiment with phrase model adaptation through training data weighting, where one assigns higher weights to relevant do- main training instances, thus causing an increase of the corresponding probabilities. As a result, translation pairs which can be obtained from rel- evant training instances will have a higher chance of being utilized during search. The main contribution of this work is design- ing several novel schemes for scoring sentences and assigning them appropriate weights to mani- fest adaptation. Our method consists of two steps: first, we find phrase alignments for the bilingual training data, then, the aligned data is scored using translation models and weights are generated. Experiments using the suggested methods and a comparison to previous work are done on two tasks: Arabic-to-English and German-to-English TED lectures translation. The results show sig- nificant improvements over the baseline, and sig- nificant improvements over previous work are re- ported when combining our suggested methods with previous work. The rest of the paper is organized as follows. Related work on adaptation and weighting is de- tailed in Section 2. The weighted …

Content truncated to 500 words