UCH-UPV English–Spanish System for WMT10

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 207–211, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics UCH-UPV English–Spanish system for WMT10 Francisco Zamora-Mart´ınez Dep. de F´ısica, Matem´aticas y Computaci´on Universidad CEU-Cardenal Herrera Alfara del Patriarca (Valencia), Spain fzamora@dsic.upv.es Germ´an Sanchis-Trilles Instituto Tecnol´ogico de Inform´atica Universidad Polit´ecnica de Valencia Valencia, Spain gsanchis@dsic.upv.es Abstract This paper describes the system developed in collabaration between UCH and UPV for the 2010 WMT. For this year’s work- shop, we present a system for English- Spanish translation. Output N-best lists were rescored via a target Neural Network Language Model, yielding improvements in the final translation quality as measured by BLEU and TER. 1 Introduction In Statistical Machine Translation (SMT), the goal is to translate a sentence f from a given source lan- guage into an equivalent sentence ˆe from a certain target language. Such statement is typically for- malised by means of the so-called log-linear mod- els (Papineni et al., 1998; Och and Ney, 2002) as follows: ˆe = argmax 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 optimised during the tuning stage with the use of a development set. Such features typically include the target lan- guage model p(e), which is one of the core com- ponents of an SMT system. In fact, most of the times it is assigned a relatively high weight in the log-linear combination described above. Tradi- tionally, language modelling techniques have been classified into two main groups, the first one in- cluding traditional grammars such as context-free grammars, and the second one comprising more statistical, corpus-based models, such as n-gram models. In order to assign a probability to a given word, such models rely on the assumption that such probability depends on the previous history, i.e. the n − 1 preceding words in the utterance. Nowadays, n-gram models have become a “de facto” standard for language modelling in state-of- the-art SMT systems. In the present work, we present a system which follows a coherent and natural evolution of prob- abilistic Language Models. Specifically, we pro- pose the use of a continuous space language model trained in the form of a Neural Network Language Model (NN LM). The use of continuous space representation of language has been successfully applied in recent NN approaches to language modelling (Bengio et al., 2003; Schwenk and Gauvain, 2002; Castro- Bleda and Prat, 2003; Schwenk et al., 2006). However, the use of Neural Network Language Models (NN LMs) (Bengio, 2008) in state-of-the- art SMT systems is not so popular. The only com- prehensive work refers to (Schwenk, 2010), where the target LM is presented in the form of a fully- connected Multilayer Perceptron. The presented system combines a standard, state-of-the-art SMT system with a NN LM via log-linear combination …

Content truncated to 500 words