Translation Model Adaptation by Resampling

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 392–399, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Translation Model Adaptation by Resampling Kashif Shah, Lo¨ıc Barrault, Holger Schwenk LIUM, University of Le Mans Le Mans, France. FirstName.LastName@lium.univ-lemans.fr Abstract The translation model of statistical ma- chine translation systems is trained on par- allel data coming from various sources and domains. These corpora are usually con- catenated, word alignments are calculated and phrases are extracted. This means that the corpora are not weighted accord- ing to their importance to the domain of the translation task. This is in contrast to the training of the language model for which well known techniques are used to weight the various sources of texts. On a smaller granularity, the automatic cal- culated word alignments differ in quality. This is usually not considered when ex- tracting phrases either. In this paper we propose a method to auto- matically weight the different corpora and alignments. This is achieved with a resam- pling technique. We report experimen- tal results for a small (IWSLT) and large (NIST) Arabic/English translation tasks. In both cases, significant improvements in the BLEU score were observed. 1 Introduction Two types of resources are needed to train statis- tical machine translation (SMT) systems: parallel corpora to train the translation model and mono- lingual texts in the target language to build the language model. The performance of both mod- els depends of course on the quality and quantity of the available resources. Today, most SMT systems are generic, i.e. the same system is used to translate texts of all kinds. Therefore, it is the domain of the training re- sources that influences the translations that are se- lected among several choices. While monolingual texts are in general easily available in many do- mains, the freely available parallel texts mainly come from international organisations, like the European Union or the United Nations. These texts, written in particular jargon, are usually much larger than in-domain bitexts. As an exam- ple we can cite the development of an NIST Ara- bic/English phrase-based translation system. The current NIST test sets are composed of a news wire part and a second part of web-style texts. For both domains, there is only a small number of in-domain bitexts available, in comparison to almost 200 millions words of out-of-domain UN texts. The later corpus is therefore likely to domi- nate the estimation of the probability distributions of the translation model. It is common practice to use a mixture language model with coefficients that are optimized on the development data, i.e. by these means on the do- main of the translation task. Domain adaptation seems to be more tricky for the translation model and it seems that very little research has been done that seeks to apply similar ideas to the translation model. To the best of our knowledge, there is no commonly accepted method to weight the bitexts coming from different sources so that the transla- tion …

Content truncated to 500 words