Machine Translation of Medical Texts in the Khresmoi Project

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 221–228, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Machine Translation of Medical Texts in the Khresmoi Project Ondˇrej Duˇsek, Jan Hajiˇc, Jaroslava Hlav´aˇcov´a, Michal Nov´ak, Pavel Pecina, Rudolf Rosa, Aleˇs Tamchyna, Zdeˇnka Ureˇsov´a, Daniel Zeman Charles University in Prague, Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics Malostransk´e n´amˇest´ı 25, 11800 Prague, Czech Republic {odusek,hajic,hlavacova,mnovak,pecina,rosa,tamchyna,uresova,zeman}@ufal.mff.cuni.cz Abstract This paper presents the participation of the Charles University team in the WMT 2014 Medical Translation Task. Our sys- tems are developed within the Khresmoi project, a large integrated project aim- ing to deliver a multi-lingual multi-modal search and access system for biomedical information and documents. Being in- volved in the organization of the Medi- cal Translation Task, our primary goal is to set up a baseline for both its subtasks (summary translation and query transla- tion) and for all translation directions. Our systems are based on the phrase- based Moses system and standard meth- ods for domain adaptation. The con- strained/unconstrained systems differ in the training data only. 1 Introduction The WMT 2014 Medical Translation Task poses an interesting challenge for Machine Translation (MT). In the “standard” translation task, the end application is the translation itself. In the Medi- cal Translation Task, the MT system is considered a part of a larger system for Cross-Lingual Infor- mation Retrieval (CLIR) and is used to solve two different problems: (i) translation of user search queries, and (ii) translation of summaries of re- trieved documents. In query translation, the end user does not even necessarily see the MT output as their queries are translated and search is performed on documents in the target language. In summary translation, the sentences to be translated come from document summaries (snippets) displayed to provide infor- mation on each of the documents retrieved by the search. Therefore, translation quality may not be the most important measure in this task – the per- formance of the CLIR system as a whole is the final criterion. Another fundamental difference from the standard task is the nature of the trans- lated texts. While we can consider document sum- maries to be ordinary texts (despite their higher in- formation density in terms of terminology from a narrow domain), search queries in the medical do- main are an extremely specific type of data, and traditional techniques for system development and domain adaptation are truly put to a test here. This work is a part of the of the large integrated EU-funded Khresmoi project.1 Among other goals, such as joint text and image retrieval of ra- diodiagnostic records, Khresmoi aims to develop technology for transparent cross-lingual search of medical sources for both professionals and laypeo- ple, with the emphasis primarily on publicly avail- able web sources. In this paper, we describe the Khresmoi sys- tems submitted to the WMT 2014 Medical Trans- lation Task. We participate in both subtasks (sum- mary translation and query translation) for all language pairs (Czech–English, …

Content truncated to 500 words