Domain Adaptation for Medical Text Translation using Web Resources

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 233–238, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Domain Adaptation for Medical Text Translation Using Web Re- sources Yi Lu, Longyue Wang, 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 takamachi660@gmail.com, vincentwang0229@hotmail.com, derekfw@umac.mo, lidiasc@umac.mo, wang2008499@gmail.com, olifran@umac.mo Abstract This paper describes adapting statistical machine translation (SMT) systems to medical domain using in-domain and general-domain data as well as web- crawled in-domain resources. In order to complement the limited in-domain corpo- ra, we apply domain focused web- crawling approaches to acquire in- domain monolingual data and bilingual lexicon from the Internet. The collected data is used for adapting the language model and translation model to boost the overall translation quality. Besides, we propose an alternative filtering approach to clean the crawled data and to further optimize the domain-specific SMT sys- tem. We attend the medical summary sentence unconstrained translation task of the Ninth Workshop on Statistical Ma- chine Translation (WMT2014). Our sys- tems achieve the second best BLEU scores for Czech-English, fourth for French-English, English-French language pairs and the third best results for re- minding pairs. 1 Introduction In this paper, we report the experiments carried out by the NLP2CT Laboratory at 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). As data in specific domain are usually rela- tively scarce, the use of web resources to com- plement the training resources provides an effec- tive way to enhance the SMT systems (Resnik and smith, 2003; Esplà-Gomis and Forcada, 2010; Pecina et al., 2011; Pecina et al., 2012; Pecina et al., 2014). In our experiments, we not only use all available training data provided by the WMT2014 standard translation task 1 (general- domain data) and medical translation task2 (in- domain data), but also acquire addition in- domain bilingual translations (i.e. dictionary) and monolingual data from online sources. First of all, we collect the medical terminolo- gies from the web. This tiny but significant par- allel data are helpful to reduce the out-of- vocabulary words (OOVs) in translation models. In addition, the use of larger language models during decoding is aided by more efficient stor- age and inference (Heafield, 2011). Thus, we crawl more in-domain monolingual data from the Internet based on domain focused web-crawling approach. In order to detect and remove out- domain data from the crawled data, we not only explore text-to-topic classifier, but also propose an alternative filtering approach combined the existing one (text-to-topic classifier) with per- plexity. After carefully pre-processing all the available training data, we apply language model adaptation and translation model adaptation us- ing various kinds of training corpora. Experi- mental results show that the presented approach- es are helpful to further boost the baseline system. The reminder of this paper …

Content truncated to 500 words