Character-Based Pivot Translation for Under-Resourced Languages and Domains

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 141–151, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Character-Based Pivot Translation for Under-Resourced Languages and Domains J¨org Tiedemann Department of Linguistics and Philology Uppsala University, Uppsala/Sweden jorg.tiedemann@lingfil.uu.se Abstract In this paper we investigate the use of character-level translation models to sup- port the translation from and to under- resourced languages and textual domains via closely related pivot languages. Our ex- periments show that these low-level models can be successful even with tiny amounts of training data. We test the approach on movie subtitles for three language pairs and legal texts for another language pair in a do- main adaptation task. Our pivot translations outperform the baselines by a large margin. 1 Introduction Data-driven approaches have been extremely suc- cessful in most areas of natural language pro- cessing (NLP) and can be considered the main paradigm in application-oriented research and de- velopment. Research in machine translation is a typical example with the dominance of statisti- cal models over the last decade. This is even en- forced due to the availability of toolboxes such as Moses (Koehn et al., 2007) which make it pos- sible to build translation engines within days or even hours for any language pair provided that ap- propriate training data is available. However, this reliance on training data is also the most severe limitation of statistical approaches. Resources in large quantities are only available for a few lan- guages and domains. In the case of SMT, the dilemma is even more apparent as parallel cor- pora are rare and usually quite sparse. Some lan- guages can be considered lucky, for example, be- cause of political situations that lead to the pro- duction of freely available translated material on a large scale. A lot of research and development would not have been possible without the Euro- pean Union and its language policies to give an example. One of the main challenges of current NLP re- search is to port data-driven techniques to under- resourced languages, which refers to the major- ity of the world’s languages. One obvious ap- proach is to create appropriate data resources even for those languages in order to enable the use of similar techniques designed for high-density lan- guages. However, this is usually too expensive and often impossible with the quantities needed. Another idea is to develop new models that can work with (much) less data but still make use of resources and techniques developed for other well-resourced languages. In this paper, we explore pivot translation tech- niques for the translation from and to resource- poor languages with the help of intermediate resource-rich languages. We explore the fact that many poorly resourced languages are closely related to well equipped languages, which en- ables low-level techniques such as character- based translation. We can show that these tech- niques can boost the performance enormously, tested for several language pairs. Furthermore, we show that pivoting can also …

Content truncated to 500 words