Transn’s Submissions to the WMT22 Translation Suggestion Task

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 1205 - 1210 December 7-8, 2022 ©2022 Association for Computational Linguistics Transn’s Submissions to the WMT22 Translation Suggestion Task Hongbao Mao Wenbo Zhang Jie Cai Transn IOL Research, Wuhan, ChinaJianwei Cheng {hubben.mao, albert01.zhang, jay.cai, nevil.cheng}@transn.com Abstract This paper describes the Transn’s submissions to the WMT2022 shared task on Translation Suggestion. Our team participated on two tasks: Naive Translation Suggestion and Translation Suggestion with Hints, focusing on two lan- guage directions Zh→En and En→Zh. Apart from the golden training data provided by the shared task, we utilized synthetic corpus to fine-tune on DeltaLM (∆LM), which is a pre- trained encoder-decoder language model. We applied two-stage training strategy on ∆LM and several effective methods to generate syn- thetic corpus, which contribute a lot to the re- sults. According to the official evaluation re- sults in terms of BLEU scores, our submis- sions in Naive Translation Suggestion En→Zh and Translation Suggestion with Hints (both Zh→En and En→Zh) ranked 1st, and Naive Translation Suggestion Zh→En also achieved comparable result to the best score. 1 Introduction Combining machine translation (MT) and human translation (HT) is becoming a popular way in translation practice, which uses a typical way of post edit (PE) – the human translators are asked to provide alternatives for the incorrect word spans in the results generated by MT (Green et al., 2013; Bahdanau et al., 2015; Vaswani et al., 2017; Zouhar et al., 2021; Yang et al., 2021). In order to improve the efficiency of PE, researchers proposed transla- tion suggestion (TS) to provide the sub-segment suggestions for the annotated incorrect word spans, and experiments show that TS can substantially reduce translators’ cognitive loads and the post- editing time (Wang et al., 2020; Lee et al., 2021; Yang et al., 2021). This paper describes the contribution of Transn IOL Research to the WMT22 Translation Sugges- tion shared task, where systems were submitted to two tasks: 1) Naive Translation Suggestion; 2) Translation Suggestion with Hints. For both tasks we trained the models on pre-trained encoder- decoder language model ∆LM (Ma et al., 2021) with the corpus which were synthesized deliber- ately, then submitted the ensemble results of the trained models. Our main contributions are: . We utilized the pre-trained language model ∆LM to generate TS, which gets good re- sults on the shared tasks, and much lower computational budget than training from raw Transformers (Vaswani et al., 2017; Junczys- Dowmunt and Grundkiewicz, 2018; Yang et al., 2021) as well as better quality. . Apart from the provided golden data anno- tated by expert translators, we proposed the constructing methods for silver and bronze data to train TS system based on parallel cor- pus and the NMT models provided by the shared tasks, which contributes a lot for the final results. . Based on the Naive Translation Suggestion models, we proposed an effective algorithm for the task of Translation Suggestion with Hints, which improves BLEU scores signifi- cantly. The rest of this paper is …

Content truncated to 500 words