Elhuyar submission to the Biomedical Translation Task 2020 on terminology and abstracts translation

Proceedings of the 5th Conference on Machine Translation (WMT), pages 813–819 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 813 Elhuyar submission to the Biomedical Translation Task 2020 on terminology and abstracts translation Ander Corral Elhuyar Foundation a.corral@elhuyar.eus Xabier Saralegi Elhuyar Foundation x.saralegi@elhuyar.eus Abstract This article describes the systems submitted by Elhuyar to the 2020 Biomedical Transla- tion Shared Task, specifically the systems pre- sented in the subtasks of terminology transla- tion for English-Basque and abstract transla- tion for English-Basque and English-Spanish. In all cases a Transformer architecture was chosen and we studied different strategies to combine open domain data with biomedical domain data for building the training corpora. For the English-Basque pair, given the scarcity of parallel corpora in the biomedical domain, we set out to create domain training data in a synthetic way. The systems presented in the terminology and abstract translation subtasks for the English-Basque language pair ranked first in their respective tasks among four par- ticipants, achieving 0.78 accuracy for terminol- ogy translation and a BLEU of 0.1279 for the translation of abstracts. In the abstract transla- tion task for the English-Spanish pair our team ranked second (BLEU=0.4498) in the case of OK sentences. 1 Introduction General purpose translation systems usually per- form poorly where domain specific knowledge is required (Koehn and Knowles, 2017). Therefore, it is essential to develop translation systems for specific domains. However, high quality parallel corpora for domain specific tasks are only avail- able for a few major languages and obtaining in domain translated data for the majority of language pairs becomes a challenge. With such scarcity of resources, various domain adaptation techniques have shown promising results (Currey et al. 2017, Saunders et al. 2019, Sennrich et al. 2017). The biomedical domain is of great interest for the application of machine translation. It is a sector of great importance in society, even more so after the advent of COVID-19, where the handling of documentary information plays a major role. There- fore, it is a scenario where machine translation can be of great help in order to facilitate the flow of information between different languages. We can find different works in the literature that address the task of developing NMT systems adapted to the biomedical domain (Yepes et al., 2017; Sennrich et al., 2017; Khan et al., 2018). In contrast, few papers focus on language pairs where training data is scarce. This is precisely the case of the task of translation from English to Basque. The work of Soto et al. (2019) is specially relevant in this case, which presents a clinical domain ori- ented system for Basque-Spanish based on RNN and Transformer architectures and which does not require bilingual domain texts but only bilingual clinical terminology (SNOMED CT). Our participation in the Biomedical Translation Shared Task addresses the translation of biomed- ical terminology from English to Basque and the translation of biomedical abstracts from English to Basque and from English to Spanish. Given the scarcity of real parallel biomedical domain corpora, …

Content truncated to 500 words