Connecting Phrase based Statistical Machine Translation Adaptation

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3135–3145, Osaka, Japan, December 11-17 2016. Connecting Phrase based Statistical Machine Translation Adaptation Rui Wang1,2, Hai Zhao1,2 ∗, Bao-Liang Lu1,2, Masao Utiyama3 ∗ and Eiichro Sumita3 1Department of Computer Science and Engineering, 2Key Lab of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong University, Shanghai, China 3National Institute of Information and Communications Technology, Kyoto, Japan wangrui.nlp@gmail.com, {zhaohai, blu}@cs.sjtu.edu.cn, {mutiyama, eiichiro.sumita}@nict.go.jp Abstract Although more additional corpora are now available for Statistical Machine Translation (SMT), only the ones which belong to the same or similar domains of the original corpus can indeed enhance SMT performance directly. A series of SMT adaptation methods have been proposed to select these similar-domain data, and most of them focus on sentence selection. In compar- ison, phrase is a smaller and more fine grained unit for data selection, therefore we propose a straightforward and efficient connecting phrase based adaptation method, which is applied to both bilingual phrase pair and monolingual n-gram adaptation. The proposed method is evalu- ated on IWSLT/NIST data sets, and the results show that phrase based SMT performances are significantly improved (up to +1.6 in comparison with phrase based SMT baseline system and +0.9 in comparison with existing methods). 1 Introduction Large corpora are important for Statistical Machine Translation (SMT) training. However only the rel- evant additional corpora, which are also called in-domain or related-domain corpora, can enhance the performance of SMT effectively. Otherwise the irrelevant additional corpora, which are also called out- of-domain corpora, may not benefit SMT (Koehn and Schroeder, 2007). SMT adaptation means selecting useful part from mix-domain (mixture of in-domain and out-of- domain) data, for SMT performance enhancement. The core task in adaptation is about how to select the useful data. Existing works have considered selection strategies with various granularities, though most of them only focus on sentence-level selection (Axelrod et al., 2011; Banerjee et al., 2012; Duh et al., 2013; Hoang and Sima’an, 2014a; Hoang and Sima’an, 2014b). There is a potential problem for sentence level adaptation: different parts of a sentence may belong to different domains. That is, it is possible that a sentence is overall out-of-domain, although part of it can be in-domain. Therefore a few works consider more granular level for selection. They build lexicon, Translation Models (TMs), reordering models or Language Models (LMs) to select fragment or directly adapt the models (Bellegarda, 2004; Deng et al., 2008; Moore and Lewis, 2010; Foster et al., 2010; Mansour and Ney, 2013; Carpuat et al., 2013; Chen et al., 2013a; Chen et al., 2013b; Sennrich et al., 2013; Mathur et al., 2014; Shi et al., 2015). One typical example of these methods is to train two Neural Network (NN) models (one from in-domain and the other from out-of-domain) and penalize the sentences/phrases similar to out-of-domain corpora (Duh et al., 2013; Joty et al., 2015; Durrani et al., 2015). As we know, Phrase Based SMT (PBSMT) mainly contains two models: translation …

Content truncated to 500 words