SRT’s Neural Machine Translation System for WMT22 Biomedical Translation Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 901 - 907 December 7-8, 2022 ©2022 Association for Computational Linguistics SRT’s Neural Machine Translation System for WMT22 Biomedical Translation Task Yoonjung Choi, Jiho Shin, Yonghyun Ryu, Sangha Kim Samsung Research, Seoul, Republic of Korea {yj0807.choi, jiho21.shin, yonghyun.ryu,sangha01.kim}@samsung.com Abstract This paper describes the Samsung Research’s Translation system (SRT) submitted to the WMT22 biomedical translation task in two language directions: and Spanish to English.English to Spanish To improve the overall quality, we adopt the deep transformer architecture and employ the back-translation strategy for monolingual corpus. One of the issues in the domain translation is to translate domain-specific terminologies well. To address this issue, we apply the soft-constrained terminology translation based on biomedical terminology dictionaries. In this paper, we provide the performance of our system with WMT20 and WMT21 biomedical testsets. Compared to the best model in WMT20 and WMT21, our system shows equal or better performance. According to the official evaluation results in terms of BLEU scores, our systems get the highest scores in both directions. 1 Introduction Neural Machine Translation (NMT) has shown rapid growth with an encoder-decoder framework, especially Transformer (Vaswani et al., 2017), in recent years. Most of the research focuses on general-purpose translation models since there are a lot of parallel data available. On the other hand, domain-specific translation, which lacks rel- atively high-quality parallel corpus available, is one of the challenges that need to be solved in the NMT task. To address this issue, there have been several approaches such as finetuning general- purpose models with in-domain data and utiliz- ing in-domain monolingual corpus through back- translation (Yeganova et al., 2021). In the domain translation, one of the issues is the terminology translation. In the case of domain- specific terms, translation results are often poor be- cause they are relatively infrequent. Yeganova et al. (2021) also mentioned that some domain-specific terms including abbreviations were not translated correctly in previous shared tasks. Moreover, when new terms are introduced such as COVID-19, it is difficult to obtain the correct translation results as they are not in the training data. To handle this issue, we adopt the soft-constrained termi- nology translation proposed by Molchanov et al. (2021), which provides the terminology constraints of the target language as input to our system with source sentences like a hint. These terminology constraints can be obtained from in-domain dictio- naries. In addition, as many domain translation stud- ies, the back-translation strategy (Sennrich et al., 2016) is applied to generate synthetic parallel data from in-domain monolingual corpus. To improve the overall performance of our system, we also employ the Deep Transformer architecture (Bapna et al., 2018) and the ensemble strategy (Sutskever et al., 2014). Moreover, to find better transla- tion results, noisy channel modeling (Yee et al., 2019) and discriminative reranking (Lee et al., 2021) are attempted. Our experiment shows that deep transformer and data augmentation by the back-translation strategy improve the overall per- formance while the performance is not improved with reranking …

Content truncated to 500 words