Target-Side Context for Discriminative Models in Statistical Machine Translation

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1704–1714, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics Target-Side Context for Discriminative Models in Statistical Machine Translation Aleˇs Tamchyna1,2, Alexander Fraser1, Ondˇrej Bojar2 and Marcin Junczys-Dowmunt3 1LMU Munich, Munich, Germany 2Charles University in Prague, Prague, Czech Republic 3Adam Mickiewicz University in Pozna´n, Pozna´n, Poland {tamchyna,bojar}@ufal.mff.cuni.cz fraser@cis.uni-muenchen.de junczys@amu.edu.pl Abstract Discriminative translation models utiliz- ing source context have been shown to help statistical machine translation perfor- mance. We propose a novel extension of this work using target context information. Surprisingly, we show that this model can be efficiently integrated directly in the de- coding process. Our approach scales to large training data sizes and results in con- sistent improvements in translation qual- ity on four language pairs. We also pro- vide an analysis comparing the strengths of the baseline source-context model with our extended source-context and target- context model and we show that our ex- tension allows us to better capture mor- phological coherence. Our work is freely available as part of Moses. 1 Introduction Discriminative lexicons address some of the core challenges of phrase-based MT (PBMT) when translating to morphologically rich languages, such as Czech, namely sense disambiguation and morphological coherence. The first issue is se- mantic: given a source word or phrase, which of its possible meanings (i.e., which stem or lemma) should we choose? Previous work has shown that this can be addressed using a discriminative lex- icon. The second issue has to do with morphol- ogy (and syntax): given that we selected the cor- rect meaning, which of its inflected surface forms is appropriate? In this work, we integrate such a model directly into the SMT decoder. This enables our classifier to extract features not only from the full source sentence but also from a limited target- side context. This allows the model to not only help with semantics but also to improve morpho- logical and syntactic coherence. For sense disambiguation, source context is the main source of information, as has been shown in previous work (Vickrey et al., 2005), (Carpuat and Wu, 2007), (Gimpel and Smith, 2008) inter alia. Consider the first set of examples in Figure 1, pro- duced by a strong baseline PBMT system. The English word “shooting” has multiple senses when translated into Czech: it may either be the act of firing a weapon or making a film. When the cue word “film” is close, the phrase-based model is able to use it in one phrase with the ambiguous “shooting”, disambiguating correctly the transla- tion. When we add a single word in between, the model fails to capture the relationship and the most frequent sense is selected instead. Wider source context information is required for correct disam- biguation. While word/phrase senses can usually be in- ferred from the source sentence, the correct se- lection of surface forms requires also information from the target. Note that we can obtain some information from the source. For example, an English subject …

Content truncated to 500 words