Document-level re-ranking with soft lexical and semantic features for statistical machine translation

Document-level Re-ranking with Soft Lexical and Semantic Features for Statistical Machine Translation Chenchen Ding tei@mibel.cs.tsukuba.ac.jp Department of Computer Science, University of Tsukuba 1-1-1 Tennodai, Tsukuba, Ibaraki, 305-8573, Japan Masao Utiyama mutiyama@nict.go.jp Eiichiro Sumita National Institute of Information and Communications Technologyeiichiro.sumita@nict.go.jp 3-5 Hikaridai, Seikacho, Sorakugun, Kyoto, 619-0289, Japan Abstract We introduce two document-level features to polish baseline sentence-level translations gen- erated by a state-of-the-art statistical machine translation (SMT) system. One feature uses the word-embedding technique to model the relation between a sentence and its context on the tar- get side; the other feature is a crisp document-level token-type ratio of target-side translations for source-side words to model the lexical consistency in translation. The weights of introduced features are tuned to optimize the sentence- and document-level metrics simultaneously on the basis of Pareto optimality. Experimental results on two different schemes with different cor- pora illustrate that the proposed approach can efficiently and stably integrate document-level information into a sentence-level SMT system. The best improvements were approximately 0.5 BLEU on test sets with statistical significance. 1 Introduction State-of-the-art statistical machine translation (SMT) systems (Koehn et al., 2007) have achieved good performance for many translations, such as French-to-English translation. The success can be attributed to the statistical model used in translation and the huge data for model training. However, the well-developed techniques of SMT are mainly focused on the sentence- level translation, i.e., the models are trained on the parallel corpus of sentence pairs, and the translation is conducted sentence-by-sentence. Because in practice sentences are usually con- tained in a document and surrounded by context, recent research has begun to focus on enhanc- ing SMT systems with the addition of document-level information. As to the features of document-level translation, a frequently discussed issue is lexical consistency in translation: i.e., words tend to be translated consistently in a document (Carpuat, 2009; Carpuat and Simard, 2012). There are also detailed discussions around the consistency of different parts of speech (Guillou, 2013; Meyer and Webber, 2013). On the basis of lexical consistency theory, many researchers focus on increasing the lexical consistency in translation (Tiedemann, 2010; Xiao et al., 2011; Ture et al., 2012). Beyond lexical consistency, there are attempts at using lexical cohesion in translation (Ben et al., 2013; Xiong et al., 2013a,b), which considers the semantic relation between words. Rather than the lexical features, the topic of Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 110 documents is also taken as a feature in some recent research (Gong et al., 2011; Eidelman et al., 2012; Xiong and Zhang, 2013; Hasler et al., 2014). Among the different features, lexical consistency is the simplest feature because it only considers the lexical words themselves. In contrast, lexical cohesion involves more semantic information, such as hypernyms and hyponyms, usually requiring a word-net. Approaches that use the document topic as a feature usually require training data, such as a document-level parallel corpus, in the training or decoding phases. In this paper, we …

Content truncated to 500 words