TermMind: Alibaba’s WMT21 Machine Translation Using Terminologies Task Submission

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 851–856 November 10–11, 2021. ©2021 Association for Computational Linguistics 851 TermMind: Alibaba’s WMT21 Machine Translation using Terminologies Task Submission Ke Wang, Shuqin Gu, Boxing Chen, Yu Zhao, Weihua Luo, Yuqi Zhang Machine Intelligence Technology Lab Alibaba Group Beijing, China {wk258730,shuqin.gsq,boxing.cbx}@alibaba-inc.com, kongyu@taobao.com,{weihua.luowh,chenwei.zyq}@alibaba-inc.com Abstract This paper describes our work in the WMT 2021 Machine Translation using Terminolo- gies Shared Task. We participate in the shared translation terminologies task in English to Chinese language pair. To satisfy terminol- ogy constraints on translation, we use a ter- minology data augmentation strategy based on Transformer model. We used tags to mark and add the term translations into the matched sentences. We created synthetic terms using phrase tables extracted from bilingual corpus to increase the proportion of term translations in training data. Detailed pre-processing and filtering on data, in-domain finetuning and en- semble method are used in our system. Our submission obtains competitive results in the terminology-targeted evaluation. 1 Introduction Terminology is important for domain-specific ma- chine translation. Each domain has its own termi- nology, which represents the important and core concepts in the domain. In the workflow of human translation, terminology is an effective method to integrate the knowledge of human translator into machine translations (Wuebker et al., 2016; Cheng et al., 2016; Álvaro Peris et al., 2017). One line of approach is “hard constraint”. The terminology is ensured to appear in the trans- lation by adding constraints in beam search de- coding (Hokamp and Liu, 2017; Post and Vilar, 2018). However, the enforcement of terminology constraints tends to reduce the fluency of transla- tion (Hasler et al., 2018), especially when there are multiple constraints or the constraint is noisy (Susanto et al., 2020). Another line of approach is “soft constraint”. Training data is augmented with placeholders or additional terminology trans- lations (Arthur et al., 2016; Song et al., 2019; Dinu et al., 2019; Chen et al., 2020; Ailem et al., 2021a). The above methods assume that the terminology translations are good ones. However, in industry or real world the terminology translations may be noisy (Li et al., 2020). And in the human trans- lation workflows the terminology constraints usu- ally need to be applied hierarchically according to priority. In these scenarios one source term will have more than one translation. Therefore, we are happy to participate in this task and develop the method to deal with 1-to-many term translations in neural machine translation systems. The structure of the paper is as follows. Sec- tion 2 describes the dataset, data pre-processing and selection. We introduce details of our system in Section 3. The experiment settings, terminolo- gies used in training and main results are intro- duced in Section 4. Finally, we conclude our work in Section 5. 2 Data 2.1 Data Source For this task, we utilize parallel data from En- glish to Chinese language provided in WMT2021: ParaCrawl v7.1, News Commentary v16, Wiki Ti- tles v3, UN Parallel Corpus V1.0, CCMT Corpus and WikiMatrix. …

Content truncated to 500 words