Phrase Training Based Adaptation for Statistical Machine Translation

Proceedings of NAACL-HLT 2013, pages 649–654, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Phrase Training Based Adaptation for Statistical Machine Translation 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 We present a novel approach for translation model (TM) adaptation using phrase train- ing. The proposed adaptation procedure is ini- tialized with a standard general-domain TM, which is then used to perform phrase training on a smaller in-domain set. This way, we bias the probabilities of the general TM towards the in-domain distribution. Experimental re- sults on two different lectures translation tasks show significant improvements of the adapted systems over the general ones. Additionally, we compare our results to mixture modeling, where we report gains when using the sug- gested phrase training adaptation method. 1 Introduction The task of domain-adaptation attempts to exploit data mainly drawn from one domain (e.g. news, parliamentary discussion) to maximize the perfor- mance on the test domain (e.g. lectures, web fo- rums). In this work, we focus on translation model (TM) adaptation. A prominent approach in recent work is weighting at different levels of granularity. Foster and Kuhn (2007) perform weighting at the corpus level, where different corpora receive differ- ent weights and are then combined using mixture modeling. A finer grained weighting is that of Mat- soukas et al. (2009), who weight each sentence in the bitexts using features of meta-information and opti- mize a mapping from the feature vectors to weights using a translation quality measure. In this work, we propose to perform TM adapta- tion using phrase training. We start from a general- domain phrase table and adapt the probabilities by training on an in-domain data. Thus, we achieve direct phrase probabilities adaptation as opposed to weighting. Foster et al. (2010) perform weighting at the phrase level, assigning each phrase pair a weight according to its relevance to the test domain. They compare phrase weighting to a “flat” model, where the weight directly approximates the phrase probability. In their experiments, the weighting method performs better than the flat model, there- fore, they conclude that retaining the original rela- tive frequency probabilities of the TM is important for good performance. The “flat” model of Foster et al. (2010) is similar to our work. We differ in the following points: (i) we use the same procedure to perform the phrase training based adaptation and the search thus avoiding inconsistencies between the two; (ii) we do not directly interpolate the original statistics with the new ones, but use a training pro- cedure to manipulate the original statistics. We per- form experiments on the publicly available IWSLT TED task, on both Arabic-to-English and German- to-English lectures translation tracks. We compare our suggested phrase training adaptation method to a variety of baselines and show its effectiveness. Fi- nally, we experiment with mixture modeling based adaptation. We compare mixture modeling to our adaptation method, and apply our method within a mixture modeling framework. …

Content truncated to 500 words