Unsupervised Adaptation for Statistical Machine Translation

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 457–465, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Unsupervised 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 In this work, we tackle the problem of language and translation models domain- adaptation without explicit bilingual in- domain training data. In such a scenario, the only information about the domain can be induced from the source-language test corpus. We explore unsupervised adaptation, where the source-language test corpus is combined with the correspond- ing hypotheses generated by the transla- tion system to perform adaptation. We compare unsupervised adaptation to su- pervised and pseudo supervised adapta- tion. Our results show that the choice of the adaptation (target) set is crucial for successful application of adaptation meth- ods. Evaluation is conducted over the German-to-English WMT newswire trans- lation task. The experiments show that the unsupervised adaptation method generates the best translation quality as well as gen- eralizes well to unseen test sets. 1 Introduction Over the last few years, large amounts of statistical machine translation (SMT) monolingual and bilin- gual corpora were collected. Early years focused on structured data translation such as newswire. Nowadays, due to the relative success of SMT, new domains of translation are being explored, such as lecture and patent translation (Cettolo et al., 2012; Goto et al., 2013). The task of domain adaptation tackles the prob- lem of utilizing existing resources mainly drawn from one domain (e.g. parliamentary discussion) to maximize the performance on the target (test) domain (e.g. newswire). To be able to perform adaptation, a target set representing the test domain is used to manipu- late the general-domain models. Previous work on SMT adaptation focused on the scenario where (small) bilingual in-domain or pseudo in-domain training data are available. Furthermore, small at- tention was given to the choice of the target set for adaptation. In this work, we explore the problem of adaptation where no explicit bilingual data from the test domain is available for training, and the only resource encapsulating information about the domain is the source-language test corpus itself. We explore how to utilize the source-language test corpus for adapting the language model (LM) and the translation model (TM). A combination of source and automatically translated target of the test set is compared to using the source side only for TM adaptation. Furthermore, we com- pare using the test set to using in-domain data and a pseudo in-domain data (e.g. news-commentary as opposed to newswire). Experiments are done on the WMT 2013 German-to-English newswire translation task. Our best adaptation method shows competitive re- sults to the best submissions of the evaluation. This paper is structured as follows. We review related work in Section 2 and introduce the basic adaptation methods in Section 3. The experimen- tal setup is described in Section 4, results are dis- cussed in Section 5 and we conclude in Section 6. …

Content truncated to 500 words