Combining Domain Adaptation Approaches for Medical Text Translation

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 254–259, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Combining Domain Adaptation Approaches for Medical Text Transla- tion Longyue Wang, Yi Lu, Derek F. Wong, Lidia S. Chao, Yiming Wang, Francisco Oliveira Natural Language Processing & Portuguese-Chinese Machine Translation Laboratory, Department of Computer and Information Science, University of Macau, Macau, China vincentwang0229@hotmail.com, {mb25435, derekfw, lidiasc, mb25433, olifran}@umac.mo Abstract This paper explores a number of simple and effective techniques to adapt statisti- cal machine translation (SMT) systems in the medical domain. Comparative exper- iments are conducted on large corpora for six language pairs. We not only compare each adapted system with the baseline, but also combine them to further improve the domain-specific systems. Finally, we attend the WMT2014 medical summary sentence translation constrained task and our systems achieve the best BLEU scores for Czech-English, English- German, French-English language pairs and the second best BLEU scores for re- minding pairs. 1. Introduction This paper presents the experiments conducted by the NLP2CT Laboratory at the University of Macau for WMT2014 medical sentence transla- tion task on six language pairs: Czech-English (cs-en), French-English (fr-en), German-English (de-en) and the reverse direction pairs, i.e., en-cs, en-fr and en-de. By comparing the medical text with common text, we discovered some interesting phenomena in medical genre. We apply domain-specific techniques in data pre-processing, language model adaptation, translation model adaptation, numeric and hyphenated words translation. Compared to the baseline systems (detailed in Section 2 & 3), the results of each method show reasonable gains. We combine individual ap- proach to further improve the performance of our systems. To validate the robustness and lan- guage-independency of individual and combined systems, we conduct experiments on the official training data (detailed in Section 3) in all six lan- guage pairs. We anticipate the numeric compari- son (BLEU scores) on these individual and com- bined domain adaptation approaches that could be valuable for others on building a real-life do- main-specific system. The reminder of this paper is organized as fol- lows. In Section 2, we detail the configurations of our experiments as well as the baseline sys- tems. Section 3 presents the specific pre- processing for medical data. In Section 4 and 5, we describe the language model (LM) and trans- lation model (TM) adaptation, respectively. Be- sides, the techniques for numeric and hyphenated words translation are reported in Section 6 and 7. Finally, the performance of design systems and the official results are reported in Section 8. 2. Experimental Setup All available training data from both WMT2014 standard translation task1 (general-domain data) and medical translation task 2 (in-domain data) are used in this study. The official medical sum- mary development sets (dev) are used for tuning and evaluating all the comparative systems. The official medical summary test sets (test) are only used in our final submitted systems. The experiments were carried out with the Moses 1.03 (Koehn et al., 2007). The translation and the re-ordering model utilizes the “grow- …

Content truncated to 500 words