CEU-UPV English-Spanish system for WMT11

Proceedings of the 6th Workshop on Statistical Machine Translation, pages 490–495, Edinburgh, Scotland, UK, July 30–31, 2011. c⃝2011 Association for Computational Linguistics CEU-UPV English–Spanish system for WMT11 Francisco Zamora-Mart´ınez D. F´ısica, Matem´atica, y Computaci´on Universidad CEU-Cardenal Herrera Alfara del Patriarca (Valencia), Spain fzamora@dsic.upv.es M.J. Castro-Bleda D. Sistemas Inform´aticos y Computaci´on Universitat Polit`ecnica de Val`encia Valencia, Spain mcastro@dsic.upv.es Abstract This paper describes the system presented for the English-Spanish translation task by the collaboration between CEU-UCH and UPV for 2011 WMT. A comparison of indepen- dent phrase-based translation models interpo- lation for each available training corpora were tested, giving an improvement of 0.4 BLEU points over the baseline. Output N-best lists were rescored via a target Neural Network Language Model. An improvement of one BLEU point over the baseline was obtained adding the two features, giving 31.5 BLEU and 57.9 TER for the primary system, com- puted over lowercased and detokenized out- puts. The system was positioned second in the final ranking. 1 Introduction The goal of Statistical Machine Translation (SMT) is to translate a sentence between two languages. Giving the source language sentence f, it would be translated to an equivalent target language sentence e. The most extended formalization is done via log- linear models (Papineni et al., 1998; Och and Ney, 2002) as follows: ˆe = arg max e K � k=1 λkhk(f, e) (1) where hk(f, e) is a score function representing an important feature for the translation of f into e, K is the number of models (or features) and λk are the weights of the log-linear combination. Typically, the weights λk are optimized during the tuning stage with the use of a development set. SMT systems rely on a bilingual sentence aligned training corpus. These sentences are aligned at the word level (Brown et al., 1993), and after that, dif- ferent hk feature functions are trained. In some prac- tical cases, the out-of-domain training data is larger than the in-domain training data. In these cases the target Language Model (LM) is composed of a lin- ear interpolation of independent LMs, one for each available training domain or corpus. Nevertheless, the training of phrase-based translation models is an open problem in these cases. Some recent works (Resnik and Smith, 2003; Ya- suda et al., ; Koehn and Schroeder, 2007; Matsoukas et al., 2009; Foster et al., 2010; Sanchis-Trilles and Casacuberta, 2010) related to corpus weight- ing, make use of data selection, data weighting, and translation model adaptation to overcome this problem. In this work, we explore a simple cor- pus weighting technique to interpolate any number of different phrase tables. Two different approaches are tested, obtaining similar performance. On the one hand, a count-based smoothing technique that applies a weight to the counting of phrases and lexi- cal links depending on the relevance of each corpus. On the other hand, a linear interpolation of indepen- dent trained phrase tables. Another important feature of this work the use of Neural Network Language Modelsis (NN LMs) (Bengio, 2008). This kind …

Content truncated to 500 words