Parallel Corpora for the Biomedical Domain

Parallel Corpora for the Biomedical Domain Aur´elie N´ev´eol†, Antonio Jimeno Yepes∗, Mariana Neves‡, Karin Verspoor⋆ † LIMSI, CNRS, Universit´e Paris-Saclay, F-91405 Orsay ∗ IBM Research Australia ‡ German Federal Institute for Risk Assessment (BfR), Germany ⋆ University of Melbourne, Australia aurelie.neveol@limsi.fr, antonio.jimeno@au1.ibm.com mariana.lara-neves@bfr.bund.de, karin.verspoor@unimelb.edu.au Abstract A vast amount of biomedical information is available in the form of scientific literature and government-authored patient information documents. While English is the most widely used language in many of these sources, there is a need to provide access to health information in languages other than English. Parallel corpora can be leveraged to implement cross-lingual information retrieval or machine translation tools. Herein, we review the extent of parallel corpus coverage in the biomedical domain. Specifically, we perform a scoping review of existing resources and we describe the recent development of new datasets for scientific literature (the EDP dataset and an extension of the Scielo corpus) and clinical trials (the ReBEC corpus). These corpora are currently being used in the biomedical task in the Conference on Machine Translation (WMT’16 and WMT’17), which illustrates their potential for improving and evaluating biomedical machine translation systems. Furthermore, we suggest additional applications for multilin- gual natural language processing using these resources, and plan to extend resource coverage to additional text genres and language pairs. Keywords: Parallel corpus, biomedical domain, multilingual applications 1. Introduction Machine translation (MT) is currently being used for a va- riety of tasks and domains. It is known to play an impor- tant role in supporting readers’ access to textual documents in a language other than their native language or for com- municating in real time. The accuracy of MT systems has improved in recent years thanks to the availability of large collections of parallel and/or comparable corpora. In turn, these resources could be leveraged by deep learning meth- ods, which created a paradigm shift for MT. MT plays an important role in the health domain. For in- stance, it has the potential to enable patients to read docu- ments written in a language in which they are not fluent and to hold a conversation with foreign health professionals in case of accidents or health issues in a foreign country. Fur- ther, it allows patients to access health information which is only available in a foreign language, for instance, in the case of disease outbreak with origin in other countries (e.g., Zika virus outbreak in Brazil (Bueno, 2017)). MT can also support researchers to access scientific liter- ature only available in a foreign language, for instance, when working on tropical diseases specific of a region or even when moving to another country for research purposes (Walker, 2016). Finally, MT can also support the biomed- ical natural language processing (BioNLP) domain when processing documents in languages other than English for which no specific NLP tools are available. This is often the case for clinical discharge reports that are usually only available in the local language. In such cases, researchers could translate the original document into English and rely …

Content truncated to 500 words