The Universitat d’Alacant hybrid machine translation system for WMT 2011

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 457–463, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics The Universitat d’Alacant hybrid machine translation system for WMT 2011 V´ıctor M. S´anchez-Cartagena, Felipe S´anchez-Mart´ınez, Juan Antonio P´erez-Ortiz Transducens Research Group Departament de Llenguatges i Sistemes Inform`atics Universitat d’Alacant, E-03071, Alacant, Spain {vmsanchez,fsanchez,japerez}@dlsi.ua.es Abstract This paper describes the machine translation (MT) system developed by the Transducens Research Group, from Universitat d’Alacant, Spain, for the WMT 2011 shared transla- tion task. We submitted a hybrid system for the Spanish–English language pair consist- ing of a phrase-based statistical MT system whose phrase table was enriched with bilin- gual phrase pairs matching transfer rules and dictionary entries from the Apertium shallow- transfer rule-based MT platform. Our hybrid system outperforms, in terms of BLEU, GTM and METEOR, a standard phrase-based statis- tical MT system trained on the same corpus, and received the second best BLEU score in the automatic evaluation. 1 Introduction This paper describes the system submitted by the Transducens Research Group (Universitat d’Alacant, Spain) to the shared translation task of the EMNLP 2011 Sixth Workshop on Statistical Machine Translation (WMT 2011). We partici- pated in the Spanish–English task with a hybrid sys- tem that combines, in a phrase-based statistical ma- chine translation (PBSMT) system, bilingual infor- mation obtained from parallel corpora in the usual way (Koehn, 2010, ch. 5), and bilingual informa- tion from the Spanish–English language pair in the Apertium (Forcada et al., 2011) rule-based machine translation (RMBT) platform. A wide range of hybrid approaches (Thurmair, 2009) may be taken in order to build a machine translation system which takes advantage of a par- allel corpus and explicit linguistic information from RBMT. In particular, our hybridisation approach di- rectly enriches the phrase table of a PBSMT system with phrase pairs generated from the explicit lin- guistic resources from an Apertium-based shallow- transfer RBMT system. Apertium, which is de- scribed in detail below, does not perform a complete syntactic analysis of the input sentences, but rather works with simpler linear intermediate representa- tions. The rest of the paper is organised as follows. Next section overviews the two MT systems we combine in our submission. Section 3 outlines related hybrid approaches, whereas our approach is described in Section 4. Sections 5 and 6 describe, respectively, the resources we used to build our submission and the results achieved for the Spanish–English lan- guage pair. The paper ends with some concluding remarks. 2 Translation approaches We briefly describe the rationale behind the PBSMT (section 2.1) and the shallow-transfer RBMT (sec- tion 2.2) systems we have used in our hybridisation approach. 2.1 Phrase-based statistical machine translation Phrase-based statistical machine translation sys- tems (Koehn et al., 2003) translate sentences by maximising the translation probability as defined by the log-linear combination of a number of fea- ture functions, whose weights are chosen to opti- 457 mise translation quality (Och, 2003). A core com- ponent of every PBSMT system is the phrase ta- ble, which contains bilingual …

Content truncated to 500 words