Context-aware Discriminative Phrase Selection for Statistical Machine Translation

Proceedings of the Second Workshop on Statistical Machine Translation, pages 159–166, Prague, June 2007. c⃝2007 Association for Computational Linguistics Context-aware Discriminative Phrase Selection for Statistical Machine Translation Jes´us Gim´enez and Llu´ıs M`arquez TALP Research Center, LSI Department Universitat Polit`ecnica de Catalunya Jordi Girona Salgado 1–3, E-08034, Barcelona {jgimenez,lluism}@lsi.upc.edu Abstract In this work we revise the application of discriminative learning to the problem of phrase selection in Statistical Machine Translation. Inspired by common tech- niques used in Word Sense Disambiguation, we train classifiers based on local context to predict possible phrase translations. Our work extends that of Vickrey et al. (2005) in two main aspects. First, we move from word translation to phrase translation. Second, we move from the ‘blank-filling’ task to the ‘full translation’ task. We report results on a set of highly frequent source phrases, obtaining a significant improvement, specially with re- spect to adequacy, according to a rigorous process of manual evaluation. 1 Introduction Translations tables in Phrase-based Statistical Ma- chine Translation (SMT) are often built on the ba- sis of Maximum-likelihood Estimation (MLE), be- ing one of the major limitations of this approach that the source sentence context in which phrases occur is completely ignored (Koehn et al., 2003). In this work, inspired by state-of-the-art Word Sense Disambiguation (WSD) techniques, we sug- gest using Discriminative Phrase Translation (DPT) models which take into account a wider feature context. Following the approach by Vickrey et al. (2005), we deal with the ‘phrase translation’ prob- lem as a classification problem. We use Support Vector Machines (SVMs) to predict phrase transla- tions in the context of the whole source sentence. We extend the work by Vickrey et al. (2005) in two main aspects. First, we move from ‘word transla- tion’ to ‘phrase translation’. Second, we move from the ‘blank-filling’ task to the ‘full translation’ task. Our approach is fully described in Section 2. We apply it to the Spanish-to-English translation of Eu- ropean Parliament Proceedings. In Section 3, prior to considering the ‘full translation’ task, we ana- lyze the impact of using DPT models for the iso- lated ‘phrase translation’ task. In spite of working on a very specific domain, a large room for improve- ment, coherent with WSD performance, and results by Vickrey et al. (2005), is predicted. Then, in Sec- tion 4, we tackle the full translation task. DPT mod- els are integrated in a ‘soft’ manner, by making them available to the decoder so they can fully interact with other models. Results using a reduced set of highly frequent source phrases show a significant improvement, according to several automatic eval- uation metrics. Interestingly, the BLEU metric (Pap- ineni et al., 2001) is not able to reflect this improve- ment. Through a rigorous process of manual eval- uation we have verified the gain. We have also ob- served that it is mainly related to adequacy. These results confirm that better phrase translation proba- bilities may be helpful for the full translation task. However, the fact that no gain in …

Content truncated to 500 words