Abu-MaTran at WMT 2014 Translation Task: Two-step Data Selection and RBMT-Style Synthetic Rules

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 171–177, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Abu-MaTran at WMT 2014 Translation Task: Two-step Data Selection and RBMT-Style Synthetic Rules Raphael Rubino⋆, Antonio Toral†, Victor M. S´anchez-Cartagena⋆‡, Jorge Ferr´andez-Tordera⋆, Sergio Ortiz-Rojas⋆, Gema Ram´ırez-S´anchez⋆, Felipe S´anchez-Mart´ınez‡, Andy Way† ⋆ Prompsit Language Engineering, S.L., Elche, Spain {rrubino,vmsanchez,jferrandez,sortiz,gramirez}@prompsit.com † NCLT, School of Computing, Dublin City University, Ireland {atoral,away}@computing.dcu.ie ‡ Dep. Llenguatges i Sistemes Inform`atics, Universitat d’Alacant, Spain fsanchez@dlsi.ua.es Abstract This paper presents the machine trans- lation systems submitted by the Abu- MaTran project to the WMT 2014 trans- lation task. The language pair concerned is English–French with a focus on French as the target language. The French to En- glish translation direction is also consid- ered, based on the word alignment com- puted in the other direction. Large lan- guage and translation models are built us- ing all the datasets provided by the shared task organisers, as well as the monolin- gual data from LDC. To build the trans- lation models, we apply a two-step data selection method based on bilingual cross- entropy difference and vocabulary satura- tion, considering each parallel corpus in- dividually. Synthetic translation rules are extracted from the development sets and used to train another translation model. We then interpolate the translation mod- els, minimising the perplexity on the de- velopment sets, to obtain our final SMT system. Our submission for the English to French translation task was ranked second amongst nine teams and a total of twenty submissions. 1 Introduction This paper presents the systems submitted by the Abu-MaTran project (runs named DCU-Prompsit- UA) to the WMT 2014 translation task for the English–French language pair. Phrase-based sta- tistical machine translation (SMT) systems were submitted, considering the two translation direc- tions, with the focus on the English to French di- rection. Language models (LMs) and translation models (TMs) are trained using all the data pro- vided by the shared task organisers, as well as the Gigaword monolingual corpora distributed by LDC. To train the LMs, monolingual corpora and the target side of the parallel corpora are first used individually to train models. Then the individ- ual models are interpolated according to perplex- ity minimisation on the development sets. To train the TMs, first a baseline is built us- ing the News Commentary parallel corpus. Sec- ond, each remaining parallel corpus is processed individually using bilingual cross-entropy differ- ence (Axelrod et al., 2011) in order to sepa- rate pseudo in-domain and out-of-domain sen- tence pairs, and filtering the pseudo out-of- domain instances with the vocabulary saturation approach (Lewis and Eetemadi, 2013). Third, synthetic translation rules are automatically ex- tracted from the development set and used to train another translation model following a novel ap- proach (S´anchez-Cartagena et al., 2014). Finally, we interpolate the four translation models (base- line, in-domain, filtered out-of-domain and rules) by minimising the perplexity obtained on the de- velopment sets and investigate the best tuning and decoding parameters. The reminder of this paper …

Content truncated to 500 words