UPM system for the translation task

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 420–425, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics UPM system for the translation task Verónica López-Ludeña Rubén San-Segundo Grupo de Tecnología del Habla Grupo de Tecnología del Habla Universidad Politécnica de Madrid Universidad Politécnica de Madrid veronicalopez@die.upm.es lapiz@die.upm.es Abstract This paper describes the UPM system for translation task at the EMNLP 2011 workshop on statistical machine translation (http://www.statmt.org/wmt11/), and it has been used for both directions: Spanish-English and English-Spanish. This system is based on Moses with two new modules for pre and post processing the sentences. The main contribution is the method proposed (based on the similarity with the source language test set) for selecting the sentences for training the models and adjusting the weights. With system, we have obtained a 23.2 BLEU for Spanish-English and 21.7 BLEU for English- Spanish. 1 Introduction The Speech Technology Group of the Universidad Politécnica de Madrid has participated in the sixth workshop on statistical machine translation in the Spanish-English and English-Spanish translation task. Our submission is based on the state-of-the-art SMT toolkit Moses (Koehn, 2010) adding a pre- processing and a post-processing module. The main contribution is a corpus selection method for training the translation models based on the similarity of each source corpus sentence with the language model of the source language test set. There are several related works on filtering the training corpus by using a similarity measure based on the alignment score or based on sentences length (Khadivi and Ney, 2005; Sanchis-Trilles et al, 2010). However, these techniques are focused on removing noisy data, i.e., their idea is to eliminate possible errors in the databases. The difference between these techniques and the method that we propose is that we do not search “bad” pairs of sentences, but we search those sentences in source training corpus that are more similar with the language model generated with the source test sentences and we select them for training. Other interesting technique of corpus selection is based on transductive learning (Ueffing, 2007). In this work, authors use of transductive semi- supervised methods for the effective use of monolingual data from the source language in order to improve translation quality. The method proposed in this paper is also applied to the validation corpus. There are other works related to select development set (Hui, 2010) that they combine different development sets in order to find the more similar one with test set. 2 Overall description of the system The translation system used is based on Moses, the software released to support the translation task (http://www.statmt.org/wmt11/) at the EMNLP 2011 workshop on statistical machine translation. Figure 1: Moses translation system 420 The phrase model has been trained following these steps (Figure 1): • Word alignment computation. GIZA++ (Och and Ney, 2003) is a statistical machine translation toolkit that is used to calculate the alignments between Spanish and English words in both direction (Spanish-English and English- Spanish). To generate the translation model, the parameter …

Content truncated to 500 words