TSMind: Alibaba and Soochow University’s Submission to the WMT22 Translation Suggestion Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1198 - 1204 December 7-8, 2022 ©2022 Association for Computational Linguistics TSMind: Alibaba and Soochow University’s Submission to the WMT22 Translation Suggestion Task Xin Ge1∗, Ke Wang1∗, Jiayi Wang1, Nini Xiao1,2, Xiangyu Duan2, Yu Zhao1, Yuqi Zhang1† 1Alibaba Group Inc. 2Soochow University {shiyi.gx,moyu.wk,joanne.wjy}@alibaba-inc.com {nnxiaonnxiao,xiangyuduan}@suda.edu.cn, {kongyu, chenwei.zyq}@alibaba-inc.com Abstract This paper describes the joint submission of Al- ibaba and Soochow University, TSMind, to the WMT 2022 Shared Task on Translation Sugges- tion (TS). We participate in the English ↔ Ger- man and English ↔ Chinese tasks. Basically, we utilize the model paradigm fine-tuning on the downstream tasks based on large-scale pre- trained models, which has recently achieved great success. We choose FAIR’s WMT19 En- glish ↔ German news translation system and MBART50 for English ↔ Chinese as our pre- trained models. Considering the task’s condi- tion of limited use of training data, we follow the data augmentation strategies proposed by Yang et al. (2021) to boost our TS model per- formance. The difference is that we further in- volve the dual conditional cross-entropy model and GPT-2 language model to filter augmented data. The leader board finally shows that our submissions are ranked first in three of four language directions in the Naive TS task of the WMT22 Translation Suggestion task. 1 Introduction Computer-aided translation (CAT) (Barrachina et al., 2009; Green et al., 2014, 2015; Knowles and Koehn, 2016) has become more and more pop- ular to help increase the quality of machine trans- lation (Lopez, 2008; Koehn, 2009) result. It also improves the efficiency of translators by combining the results of machine translation and the content edited by translators in the process of translation or post-editing (Bowker, 2002; Lengyel and Ugray, 2004; Bowker and Fisher, 2010; Bowker, 2014; Chatterjee, 2019). Post-editing based on machine translation is typ- ical in CAT. Recent works (Domingo et al., 2016; González-Rubio et al., 2016; Peris et al., 2017) propose interactive protocols and algorithms so that humans and machines can collaborate during ∗indicates equal contribution. †indicates the corresponding author. translation, and machines can automatically pro- vide feedback on humans’ edits. One interesting mode is Translation Suggestion (TS) (Yang et al., 2021), which offers alternatives for specific spans of words in the generated machine translation. It will be convenient if the model refines translation results in those specified locations with potential translation errors. Yang et al. (2021) released a benchmark dataset for TS, WeTS, which is one of the shared tasks in WMT22. At the same time, they proposed an end-to-end Transformer-like model for TS as the benchmark system. However, the lack of many labeled TS data limits the training of a large Transformer model to some extent. Though Yang et al. (2021) have tried to uti- lize XLM-Roberta (Conneau et al., 2019) to initial- ize the encoder of the Transformer, the decoder has to be trained from scratch, which leads to relatively low BLEU scores for some specific TS spans. We investigate the potential of other encoder-decoder …

Content truncated to 500 words