Multi-domain Adaptation for Statistical Machine Translation Based on Feature Augmentation

Multi-domain Adaptation for Statistical Machine Translation Based on Feature Augmentation Kenji Imamura kenji.imamura@nict.go.jp Eiichiro Sumita National Institute of Information and Communications Technology,eiichiro.sumita@nict.go.jp 3-5 Hikaridai, Seika-cho, Soraku-gun, Kyoto 619-0289, Japan Abstract Domain adaptation is a major challenge when applying machine translation to practical tasks. In this paper, we present domain adaptation methods for machine translation that assume mul- tiple domains. The proposed methods combine two model types: a corpus-concatenated model covering multiple domains and single-domain models that are accurate but sparse in specific domains. We combine the advantages of both models using feature augmentation for domain adaptation in machine learning. Our experimental results show that the BLEU scores of the proposed method clearly sur- pass those of single-domain models for low-resource domains. For high-resource domains, the scores of the proposed method were superior to those of both single-domain and corpus- concatenated models. Even in domains having a million bilingual sentences, the translation quality was at least preserved and even improved in some domains. These results demonstrate that state-of-the-art domain adaptation can be realized with appropriate settings, even when using standard log-linear models. 1 Introduction Machine translation is used for translating a variety of text types, including speech. However, it remains challenging to appropriately translate texts across all domains and only a limited number of domains have been targeted. The most promising approach to improve translation quality is to train the translator on massive bilingual corpora. However, collecting such corpora is challenging and expensive in several domains. Domain adaptation, which improves target domain quality by using data from another domain, has been proposed as a solution (Foster and Kuhn, 2007; Foster et al., 2010; Axelrod et al., 2011; Bisazza et al., 2011; Sennrich, 2012; Sennrich et al., 2013). This technique is important when applying machine translation to practical tasks. This paper presents methods of domain adaptation for statistical machine translation (SMT) that assume multiple domains. The proposed methods combine multiple models using log-linear interpolation. These are simple yet effective approaches to take advantage of multiple domains based on feature augmentation (Daum´e, 2007), a domain adaptation technique used in machine learning. We propose the following two methods. 1. Simultaneous optimization of multiple domains: this method uses an optimizer extended to multiple domains to optimize an augmented feature space. 2. Optimization of one domain at a time: this method restricts the feature space and regards ������������������������������������������������������ ���������������������������������������� this space as that used in the standard log-linear model. This can be realized via a slight modification of existing translation systems. Both methods use a corpus-concatenated model, which covers multiple domains and con- tains few unknown words, and single-domain models, which are accurate in their specific do- mains. In addition, we tune the hyper-parameter of the multiple-model combination. With ap- propriate settings, state-of-the-art domain adaptation can be realized even when using standard log-linear models. In this study, we use phrase-based statistical machine translation (PBSMT) (Koehn et al., 2003, 2007) with preordering. The remainder of this paper is organized as follows. Section 2 briefly reviews domain adaptation …

Content truncated to 500 words