Improving Neural Machine Translation Formality Control with Domain Adaptation and Reranking-based Transductive Learning

Proceedings of the 20th International Conference on Spoken Language Translation (IWSLT 2023), pages 180–186 July 13-14, 2023 c⃝2023 Association for Computational Linguistics Improving Neural Machine Translation Formality Control with Domain Adaptation and Reranking-based Transductive Learning Zhanglin Wu, Zongyao Li, Daimeng Wei, Hengchao Shang, Jiaxin Guo, Xiaoyu Chen, Zhiqiang Rao, Zhengzhe Yu, Jinlong Yang, Shaojun Li, Yuhao Xie, Bin Wei, Jiawei Zheng, Ming Zhu, Lizhi Lei, Hao Yang, Yanfei Jiang Huawei Translation Service Center, Beijing, China {wuzhanglin2,lizongyao,weidaimeng,shanghengchao,guojiaxin1,chenxiaoyu35, raozhiqiang,yuzhengzhe,yangjinlong7,lishaojun18,xieyuhao2,weibin29, zhengjiawei15,zhuming47,leilizhi,yanghao30,jiangyanfei}@huawei.com Abstract This paper presents Huawei Translation Ser- vice Center (HW-TSC)’s submission on the IWSLT 2023 formality control task, which pro- vides two training scenarios: supervised and zero-shot, each containing two language pairs, and sets constrained and unconstrained condi- tions. We train the formality control models for these four language pairs under these two conditions respectively, and submit the corre- sponding translation results. Our efforts are di- vided into two fronts: enhancing general trans- lation quality and improving formality control capability. According to the different require- ments of the formality control task, we use a multi-stage pre-training method to train a bilin- gual or multilingual neural machine translation (NMT) model as the basic model, which can im- prove the general translation quality of the base model to a relatively high level. Then, under the premise of affecting the general translation quality of the basic model as little as possi- ble, we adopt domain adaptation and reranking- based transductive learning methods to improve the formality control capability of the model. 1 Introduction Machine translation (MT) (Lopez, 2008; Vaswani et al., 2017) models typically return one single translation for each input sentence. This means that when the input sentence is ambiguous, the MT model must choose a translation from among var- ious valid options, without regard to the intended use case or target audience. Therefore, there is a need to control certain attributes (Schioppa et al., 2021) of the text generated in a target language such as politeness (Sennrich et al., 2016a; Feely et al., 2019) or formality (Niu et al., 2017, 2018; Viswanathan et al., 2020). The lack of gold translation with alternate for- mality for supervised training and evaluation has lead researchers to rely on synthetic supervision training and manual evaluation in past work (Niu and Carpuat, 2020). Fortunately, the IWSLT for- mality control task now provides a new benchmark1 (N˘adejde et al., 2022; Agarwal et al., 2023) by contributing high-quality training datasets and test datasets for multiple language pairs. This paper presents HW-TSC’s submission on the IWSLT 2023 formality control task. How for- mality distinctions are expressed grammatically and lexically can vary widely by language. Thus, we participate in the formality control task of all these four language pairs to investigate a general formality control method that can be applied to different language pair. In addition, we also inves- tigate the difference in formality control between constrained and unconstrained conditions by intro- ducing the mBART model (Liu et al., 2020) under unconstrained condition. 2 Data 2.1 Pre-training Data We use the CCMatrix2 …

Content truncated to 500 words