The ADAPT’s Submissions to the WMT20 Biomedical Translation Task

Proceedings of the 5th Conference on Machine Translation (WMT), pages 841–848 Online, November 19–20, 2020. c⃝2020 Association for Computational Linguistics 841 The ADAPT’s Submissions to the WMT20 Biomedical Translation Task Prashanth Nayak, Rejwanul Haque and Andy Way The ADAPT Centre, School of Computing Dublin City University, Dublin, Ireland firstname.lastname@adaptcentre.ie Abstract This paper describes the ADAPT Centre’s sub- missions to the WMT20 Biomedical Transla- tion Shared Task for English-to-Basque. We present the machine translation (MT) systems that were built to translate scientific abstracts and terms from biomedical terminologies, and using the state-of-the-art neural MT (NMT) model: Transformer. In order to improve our baseline NMT system, we employ a number of methods, e.g. “pseudo” parallel data selection, monolingual data selection for synthetic cor- pus creation, mining monolingual sentences for adapting our NMT systems to this task, hy- perparameters search for Transformer in low- resource scenarios. Our experiments show that systematic addition of the aforementioned techniques to the baseline yields an excellent performance in the English-to-Basque transla- tion task. 1 Introduction The ADAPT Centre participated in the Biomedical Translation Shared Task of the Fifth Conference of Machine Translation (WMT20). This task is about evaluating systems on the translation of documents from the biomedical domain. The test data consists of biomedical abstracts and terminologies. The task addresses a number of language pairs, and we par- ticipated in the English-to-Basque translation task. To make the readers familiar with the biomedical translation task and to understand the challenges of this task, we show a couple of examples from the blind test set and two terminological expressions from terminology test set in Table 1. For building our MT systems we used the Trans- former model (Vaswani et al., 2017). Our strategies to build the competitive MT systems for the task roughly include (i) pseudo in-domain parallel and monolingual data selection, (ii) augmenting train- ing data (Sennrich et al., 2016a; Zhang and Zong, 2016; Burlot and Yvon, 2018; Poncelas et al., 2018; Caswell et al., 2019; Chen et al., 2019), (iii) mining (1) No cardiovascular risk factor differences were found in terms of age. (2) Congenital tumors show a different pattern than tumors in other pediatric ages. (3) Open bite of thyroid gland, sequela (4) poisoning by oxytocic drugs, undetermined, subsequent encounter Table 1: Sentences ((1) and (2)) from the blind test set and sample terminological expressions ((3) and (4)). monolingual sentences to adapt our NMT systems to the task, and (iv) finding the optimal hyperpa- rameter configuration for Transformer in this low- resource settings. The remainder of the paper is organized as fol- lows. In Section 2, we present our methods, and Section 3 details of the data sets used. Section 4 presents the results and discussions, while Section 5 concludes our work with avenues for future work. 2 Our Approaches 2.1 Selecting pseudo In-domain Parallel Sentences The shared task organisers released parallel training data with a limited number of in-domain examples (only 24,247). The organisers also provided the par- ticipants with moderate-sized three out-of-domain corpora (totalling …

Content truncated to 500 words