LUL’s WMT22 Automatic Post-Editing Shared Task Submission

Proceedings of the Seventh Conference on Machine Translation (WMT), pages 689 - 693 December 7-8, 2022 ©2022 Association for Computational Linguistics LUL’s WMT22 Automatic Post-Editing Shared Task Submission Xiaoying Huang1,2 and Xingrui Lou1,2 and Fan Zhang1 and Mei Tu1 1Samsung Research China – Beijing (SRC-B) 2State Key Laboratory of Media Convergence and Communication, Communication University of China {huenhoying, louxingrui}@cuc.edu.cn, {zhang.fan, mei.tu}@samsung.com Abstract By learning the human post-edits, the automatic post-editing (APE) models are often used to modify the output of the machine translation (MT) system to make it as close as possible to human translation. We introduce the sys- tem used in our submission of WMT’22 Au- tomatic Post-Editing (APE) English-Marathi (En-Mr) shared task. In this task, we first train the MT system of En-Mr to generate additional machine-translation sentences. Then we use the additional triple to bulid our APE model and use APE dataset to further fine-tuning. In- spired by the mixture of experts (MoE), we use GMM algorithm to roughly divide the text of APE dataset into three categories. After that, the experts are added to the APE model and dif- ferent domain data are sent to different experts. Finally, we ensemble the models to get better performance. Our APE system significantly im- proves the translations of provided MT results by -2.848 and +3.74 on the development dataset in terms of TER and BLEU, respectively. Fi- nally, the TER and BLEU scores are improved by -1.22 and +2.41 respectively on the blind test set.1 1 Introduction Automatic Post-Editing (APE) is the task of au- tomatically editing the translations of MT system. By using APE models, we can transfer MT system from general domain to specific domain and then reduce the workload of human post-edits. WMT has been holding APE task competitions in dif- ferent languages and fields since 2015. Now the APE models are often based on transformer and improved on this basis. WMT 2022’s Automatic Post-Editing task fo- cused on English-Marathi language pairs. The difference from the previous competition is that the target language is changed to Marathi. Be- sides, two new fields, medical care and tourism, 1Work performed during internship in Samsung Research China - Beijing are added. Participants are provided a training set with 18000 instances, a development set and a test set with 1000 instances respectively. Each dataset consists of source, machine-translation and post- edit triplets. The source sentences in English come from the healthcare, tourism, and general/news do- mains. The MT outputs are automatic translations to Marathi. The post-edits are human revisions of the target elements. This synthetic training data is prepared as a part of the 2022 APE shared task. The data is created by taking a parallel corpus, where the source data is translated using an MT system, and the references are considered as post-edits. Par- ticipants are also allowed to use any additional data for systems training. Last year’s research mainly focused on trans- fer learning and data augmentation. Sharma et al. (2021) utilizes the most advanced En-De machine translation …

Content truncated to 500 words