HW-TSC’s Submissions to the WMT22 Word-Level Auto Completion Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1192 - 1197 December 7-8, 2022 ©2022 Association for Computational Linguistics HW-TSC’s Submissions to the WMT22 Word-Level Auto Completion Task Hao Yang, Hengchao Shang, Zongyao Li, Daimeng Wei, Xianghui He, Xiaoyu Chen, Zhengzhe Yu, Jiaxin Guo, Jinlong Yang Shaojun Li, Yuanchang Luo, Yuhao Xie, Lizhi Lei, Ying Qin Huawei Translation Service Center, Beijing, China {yanghao30, shanghengchao,lizongyao,weidaimeng,hexianghui, chenxiaoyu35,yuzhengzhe,guojiaxin1,yangjinlong7, lishaojun18,luoyuanchang,xieyuhao2,leilizhi,qinying}@huawei.com Abstract This paper presents the submissions of Huawei Translation Services Center (HW-TSC) to WMT 2022 Word-Level AutoCompletion Task. We propose an end-to-end autoregressive model with bi-context based on Transformer to solve the current task. The model uses a mix- ture of subword and character encoding units to realize the joint encoding of human input, the context of the target side and the decoded se- quence, which ensures full utilization of infor- mation. We use one model to solve four types of data structures in the task. During training, we try using a machine translation model as the pre-trained model and fine-tune it for the task. We also add BERT-style MLM data at the fine-tuning stage to improve model perfor- mance. We participate in zh→en, en→de, and de→en directions and win the first place in all the three tracks. Particularly, we outperform the second place by more than 5% in terms of accuracy on the zh→en and en→de tracks. The result is buttressed by human evaluations as well, demonstrating the effectiveness of our model. 1 Introduction In recent years, machine translation quality has improved significantly with advances in model ar- chitecture (Sutskever et al., 2014; Bahdanau et al., 2014; Gehring et al., 2017; Vaswani et al., 2017), bilingual data availability, as well as data augmenta- tion strategies (Liu et al., 2016; Freitag et al., 2017; Johnson et al., 2017; Zhang et al., 2018; Edunov et al., 2018; Wu et al., 2019; Li et al., 2019). In scenarios where machine translation is used to fa- cilitate understanding, machine translation outputs can basically satisfy audience’s demands. However, in areas where translation quality is crucial (such as translating product manual, patent description, etc.), post-editing is required. Techniques to im- prove post-editing efficiency are meaningful and necessary. Researches (Barrachina et al., 2009; Green et al., 2014; Knowles and Koehn, 2016; Santy et al., 2019) in this regard falls into this cate- gory of computer-aided translation (CAT). Word-Level auto Completion, as a new task in WMT22, fall into this category as well. This task aims at auto-completing a target word given a source sentence, translation context, and a human- typed character sequence, so as to improve post- editing efficiency. Li et al. (2021) define the task in detail, offer comprehensive analysis and provide a baseline system. For this task, we choose the subword-level mod- eling strategy (Kudo and Richardson, 2018). Com- paring with word-level modeling, this strategy en- ables the usage of pre-trained models from other mainstream tasks, and solves the out-of-vocabulary (OOV) issues at the same time. As human-typed input is just several characters of the target …

Content truncated to 500 words