HW-TSC Translation Systems for the WMT22 Chat Translation Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 962 - 968 December 7-8, 2022 ©2022 Association for Computational Linguistics HW-TSC Translation Systems for the WMT22 Chat Translation Task Jinlong Yang, Zongyao Li, Daimeng Wei, Hengchao Shang, Xiaoyu Chen, Zhengzhe Yu, Zhiqiang Rao, Shaojun Li, Zhanglin Wu, Yuhao Xie, Yuanchang Luo, Ting Zhu, Yanqing Zhao, Lizhi Lei, Hao Yang, Ying Qin Huawei Translation Service Center, Beijing, China {yangjinlong7,lizongyao,weidaimeng,shanghengchao,chenxiaoyu35, yuzhengzhe,raozhiqiang,lishaojun18,wuzhanglin2,xieyuhao2,luoyuanchang, zhuting20,zhaoyanqing,leilizhi,yanghao30,qinying}@huawei.com Abstract This paper describes the submissions of Huawei Translation Services Center(HW-TSC) to WMT22 chat translation shared task on English↔German (en-de) bidirection with re- sults of zero-shot and few-shot tracks. We use the deep transformer architecture with a larger parameter size. Our submissions to the WMT21 News Translation task are used as the baselines. We adopt strategies such as back translation, forward translation, domain trans- fer, data selection, and noisy forward transla- tion in task, and achieve competitive results on the development set. We also test the effec- tiveness of document translation on chat tasks. Due to the lack of chat data, the results on the development set show that it is not as effective as sentence-level translation models. 1 Introduction Neural Machine Translation (NMT) (Sutskever et al., 2014; Bahdanau et al., 2015; Gehring et al., 2017) has achieved good translation results in most scenarios, but few researches have been done in the field of chat translation, mainly because of insuffi- cient chat data. WMT20 holds the chat translation shared task (Farajian et al., 2020) for the first time. The data set mainly includes pre-sales conversations between customers and agents (meal booking, air ticket reservation, etc.). This year, the data set focuses on post-sales conversations between customers and agents. Although the translation content is all about chat, the domains are slightly different. The results show that the data from previous years can not ef- fectively improve the quality of the model for this year’s task. We participate in the en-de bidirectional trans- lation task. The en-de bidirectional models we submitted to the WMT21 news task (Wei et al., 2021) are used as the baseline models and the ar- chitecture is deep transformer (Vaswani et al., 2017; Dou et al., 2018). Commonly-used optimization strategies are used, such as domain transfer, data selection, back translation, self-training, noisy self- training, finetuning and model averaging. Considering that the chat task is a context-aware translation task, we conduct a series of document- level (Wang et al., 2017) experiments using WMT document data, but it does not work well on devel- opment sets. The analysis shows that the document data deviates greatly from the chat domain, and the data therefore cannot effectively improve chat translation quality. According to the results, the best models are obtained by selecting in-domain data from out domain data by the development sets. This paper is structured as follows: Section 2 describes our data volume and data pre-processing method. The model structure and method we used are presented in Section 3. Section 4 details our experiment setting. We present the results in …

Content truncated to 500 words