UPM system for WMT 2012

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 338–344, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics UPM system for WMT 2012 Verónica López-Ludeña, Rubén San-Segundo and Juan M. Montero GTH-IEL-ETSI Telecomunicación Universidad Politécnica de Madrid {veronicalopez, lapiz, juancho}@die.upm.es Abstract This paper describes the UPM system for the Spanish-English translation task at the NAACL 2012 workshop on statistical ma- chine translation. This system is based on Mo- ses. We have used all available free corpora, cleaning and deleting some repetitions. In this paper, we also propose a technique for select- ing the sentences for tuning the system. This technique is based on the similarity with the sentences to translate. With our approach, we improve the BLEU score from 28.37% to 28.57%. And as a result of the WMT12 chal- lenge we have obtained a 31.80% BLEU with the 2012 test set. Finally, we explain different experiments that we have carried out after the competition. 1 Introduction The Speech Technology Group at the Technical University of Madrid has participated in the sev- enth workshop on statistical machine translation in the Spanish-English translation task. Our submission is based on the state-of-the-art SMT toolkit Moses (Koehn et al., 2007). Firstly, we have proved different corpora for training the system: cleaning the whole corpus and deleting some repetitions in order to have a better perfor- mance of the translation model. There are several related works on filtering the training corpus by removing noisy data that use a similarity measure based on the alignment score or based on sentences length (Khadivi and Ney, 2005). In this paper, we also propose a technique for selecting the most appropriate sentences for tuning the system, based on the similarity with the Span- ish sentences to translate. This technique is an up- date of the technique proposed by our group in the last WMT11 challenge (López-Ludeña and San- Segundo, 2011). There are other works related to select the development set (Hui et al., 2010) that combine different development sets in order to find the more similar one with test set. There are also works related to select sentences, but for training instead of tuning, based on the sim- ilarity with the source test sentences. Some of them are based on transductive learning: semi- supervised methods for the effective use of mono- lingual data from the source language in order to improve translation quality (Ueffing, 2007); meth- ods using instance selection with feature decay algorithms (Bicici and Yuret, 2011); or using TF- IDF algorithm (Lü et al., 2007). There are also works based on selecting training material with active learning: using language model adaptation (Shinozaki et al., 2011); or perplexity-based meth- ods (Mandal et al., 2008). In this work, we have used the proposed selec- tion method only for tuning. The rest of the paper is organized as follows. Next section overviews the system. Section 3 de- scribes the used corpora. Section 4 explains the experiments carried out before the competition. Section 5 describes the sentences …

Content truncated to 500 words