DCU Terminology Translation System for Medical Query Subtask at WMT14

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 239–245, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics The DCU Terminology Translation System for the Medical Query Subtask at WMT14 Tsuyoshi Okita, Ali Hosseinzadeh Vahid, Andy Way, Qun Liu Dublin City University, School of Computing Glasnevin, Dublin 9 Ireland {tokita,avahid,away,qliu}@computing.dcu.ie Abstract This paper describes the Dublin City University terminology translation system used for our participation in the query translation subtask in the medical trans- lation task in the Workshop on Statisti- cal Machine Translation (WMT14). We deployed six different kinds of terminol- ogy extraction methods, and participated in three different tasks: FR–EN and EN– FR query tasks, and the CLIR task. We obtained 36.2 BLEU points absolute for FR–EN and 28.8 BLEU points absolute for EN–FR tasks where we obtained the first place in both tasks. We obtained 51.8 BLEU points absolute for the CLIR task. 1 Introduction This paper describes the terminology translation system developed at Dublin City University for our participation in the query translation subtask at the Workshop on Statistical Machine Translation (WMT14). We developed six kinds of terminol- ogy extraction methods for the problem of medi- cal terminology translation, especially where rare and new words are considered. We have several motivations which we address before providing a description of the actual algorithms undeprinning our work. First, terminology translation cannot be seen just as a simple extension of the translation process if we use an analogy from human translation. Ter- minology translation can be considered as more important and a quite different task than transla- tion per se, so we need a considerably different way of solving this particular problem. Bilingual terminology selection has been claimed to be the touchstone in human translation, especially where scientific and legal translation are concerned. Ter- minology selection is often the hardest and most time-consuming process in the translation work- flow. Depending on the particular requirements of the use-case (Way, 2013), users may not object to disfluent translations, but will invariably be very sensitive to the wrong selection of terminology, even if the meaning of the chosen terms is correct. This is especially true if this selected terminology does not match with that preferred by the users themselves, in which case users are likely to ex- press some kind of complaint; it may even be that the entire translation is rejected as sub-standard or inappropriate on such grounds. Second, we look at how to handle new and rare words. If we inspect the process of human trans- lation more closely, it is easy to identify several differences compared to the methods used in sta- tistical MT (SMT). Unless stipulated by the client, the selection of bilingual terminology can be a highly subjective process. Accordingly, it is not necessarily the bilingual term-pair with the highest probability that is chosen by the human translator. It is often the case that statistical methods often forget about or delete less frequent n-grams, but rely on more frequent n-grams using maximum likelihood …

Content truncated to 500 words