UPV-PRHLT English–Spanish System for WMT10

Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pages 172–176, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics UPV-PRHLT English–Spanish system for WMT10 Germ´an Sanchis-Trilles and Jes´us Andr´es-Ferrer and Guillem Gasc´o Jes´us Gonz´alez-Rubio and Pascual Mart´ınez-G´omez and Martha-Alicia Rocha Joan-Andreu S´anchez and Francisco Casacuberta Instituto Tecnol´ogico de Inform´atica Departamento de Sistemas Inform´aticos y Computaci´on Universidad Polit´ecnica de Valencia {gsanchis|jandres|fcn}@dsic.upv.es {ggasco|jegonzalez|pmartinez}@dsic.upv.es {mrocha|jandreu}@dsic.upv.es Abstract In this paper, the system submitted by the PRHLT group for the Fifth Work- shop on Statistical Machine Translation of ACL2010 is presented. On this evalua- tion campaign, we have worked on the English–Spanish language pair, putting special emphasis on two problems derived from the large amount of data available. The first one, how to optimize the use of the monolingual data within the language model, and the second one, how to make good use of all the bilingual data provided without making use of unnecessary com- putational resources. 1 Introduction For this year’s translation shared task, the Pat- tern Recognition and Human Language Technolo- gies (PRHLT) research group of the Universidad Polit´ecnica de Valencia submitted runs for the English–Spanish translation task. In this paper, we report the configuration of such a system, together with preliminary experiments performed to estab- lish the final setup. As in 2009, the central focus of the Shared Task is on Domain Adaptation, where a system typi- cally trained using out-of-domain data is adjusted to translate news commentaries. For the preliminary experiments, we used only a small amount of the largest available bilingual cor- pus, i.e. the United Nations corpus, by including into our system only those sentences which were considered similar. Language model interpolation using a develop- ment set was explored in this work, together with a technique to cope with the problem of ”out of vocabulary words”. Finally, a reordering constraint using walls and zones was used in order to improve the perfor- mance of the submitted system. In the final evaluation, our system was ranked fifth, considering only primary runs. 2 Language Model interpolation Nowadays, it is quite common to have very large amounts of monolingual data available from sev- eral different domains. Despite of this fact, in most of the cases we are only interested in trans- lating from one specific domain, as is the case in this year’s shared task, where the provided mono- lingual training data belonged to European parlia- mentary proceedings, news related domains, and the United Nations corpus, which consists of data crawled from the web. Although the most obvious thing to do is to con- catenate all the data available and train a single language model on the whole data, we also inves- tigated a “smarter” use of such data, by training one language model for each of the available cor- pora. 3 Similar sentences selection Currently, it is common to of huge bilingual cor- pora for SMT. For some common language pairs, corpora of millions of parallel sentences are avail- able. In some of the cases big …

Content truncated to 500 words