Adaptation of Reordering Models for Statistical Machine Translation

Proceedings of NAACL-HLT 2013, pages 938–946, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Adaptation of Reordering Models for Statistical Machine Translation Boxing Chen, George Foster and Roland Kuhn National Research Council Canada first.last@nrc-cnrc.gc.ca Abstract Previous research on domain adaptation (DA) for statistical machine translation (SMT) has mainly focused on the translation model (TM) and the language model (LM). To the best of our knowledge, there is no previous work on reordering model (RM) adaptation for phrase- based SMT. In this paper, we demonstrate that mixture model adaptation of a lexical- ized RM can significantly improve SMT per- formance, even when the system already con- tains a domain-adapted TM and LM. We find that, surprisingly, different training corpora can vary widely in their reordering character- istics for particular phrase pairs. Furthermore, particular training corpora may be highly suit- able for training the TM or the LM, but unsuit- able for training the RM, or vice versa, so mix- ture weights for these models should be esti- mated separately. An additional contribution of the paper is to propose two improvements to mixture model adaptation: smoothing the in-domain sample, and weighting instances by document frequency. Applied to mixture RMs in our experiments, these techniques (es- pecially smoothing) yield significant perfor- mance improvements. 1 Introduction A phrase-based statistical machine translation (SMT) system typically has three main components: a translation model (TM) that contains information about how to translate word sequences (phrases) from the source language to the target language, a language model (LM) that contains information about probable word sequences in the target lan- guage, and a reordering model (RM) that indicates how the order of words in the source sentence is likely to influence the order of words in the target sentence. The TM and the RM are trained on parallel data, and the LM is trained on target-language data. Usage of language and therefore the best translation practice differs widely across genres, topics, and di- alects, and even depends on a particular author’s or publication’s style; the word “domain” is often used to indicate a particular combination of all these fac- tors. Unless there is a perfect match between the training data domain and the (test) domain in which the SMT system will be used, one can often get bet- ter performance by adapting the system to the test domain. In offline domain adaptation, the system is pro- vided with a sample of translated sentences from the test domain prior to deployment. In a popular variant of offline adaptation, linear mixture model adaptation, each training corpus is used to gener- ate a separate model component that forms part of a linear combination, and the sample is used to as- sign a weight to each component (Foster and Kuhn, 2007). If the sample resembles some of the corpora more than others, those corpora will receive higher weights in the combination. Previous research on domain adaptation for SMT has focused on the TM and the LM. Such research is easily motivated: …

Content truncated to 500 words