Local lexical adaptation in Machine Translation through triangulation: SMT helping SMT

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 232–240, Beijing, August 2010 Local lexical adaptation in Machine Translation through triangulation: SMT helping SMT Josep Maria Crego LIMSI-CNRS jmcrego@limsi.fr Aurélien Max LIMSI-CNRS Univ. Paris Sud amax@limsi.fr François Yvon LIMSI-CNRS Univ. Paris Sud yvon@limsi.fr Abstract We present a framework where auxiliary MT systems are used to provide lexical predictions to a main SMT system. In this work, predictions are obtained by means of pivoting via auxiliary languages, and introduced into the main SMT sys- tem in the form of a low order language model, which is estimated on a sentence- by-sentence basis. The linear combination of models implemented by the decoder is thus extended with this additional lan- guage model. Experiments are carried out over three different translation tasks using the European Parliament corpus. For each task, nine additional languages are used as auxiliary languages to obtain the trian- gulated predictions. Translation accuracy results show that improvements in trans- lation quality are obtained, even for large data conditions. 1 Introduction Important improvements are yet to come regard- ing the performance of Statistical Machine Trans- lation systems. Dependence on training data and limited modelling expressiveness are the focus of many research efforts, such as using monolingual corpora for the former and syntactic models for the latter. Another promising approach consists in ex- ploiting complementary sources of information in order to build better translations, as done by consensus-based system combination (e.g. (Ma- tusov et al., 2008)). This, however, requires to have several systems available for the same lan- guage pair. Considering that the same training data would be available to all systems, differences in translation modelling are expected to produce redundant and complementary hypotheses. Mul- tisource translation (e.g. (Och and Ney, 2001; Schwartz, 2008)) is a variant, involving source texts available in several languages which can be translated by systems for different language pairs and whose outputs can be successfully combined into better translations (Schroeder et al., 2009). One theoretical expectation of multisource trans- lation is that it can successfully reduce ambiguity of the original source text, but does so under the rare conditions of availability of existing (accu- rate) translations. In contrast, pivot-based system combination (e.g. (Utiyama and Isahara, 2007; Wu and Wang, 2007)) aims at compensating the lack of training data for a given language pair by producing translation hypotheses obtained by piv- oting via an intermediary language for which bet- ter systems are available. These techniques generally produce a search space that differs from that of the direct transla- tion systems. As such, they create a new transla- tion system out of various systems for which di- agnosis becomes more difficult. This paper instead focusses on improving a sin- gle system, which should be state-of-the-art as regards data and models. We propose a frame- work in which information coming from external sources is used to boost lexical choices and guide the decoder into making more informed choices.1 1We performed initial experiments where the comple- mentary information was …

Content truncated to 500 words