Netmarble AI Center’s WMT21 Automatic Post-Editing Shared Task Submission

Proceedings of the Sixth Conference on Machine Translation (WMT), pages 307–314 November 10–11, 2021. ©2021 Association for Computational Linguistics 307 Netmarble AI Center’s WMT21 Automatic Post-Editing Shared Task Submission Shinhyeok Oh∗, Sion Jang*, Hu Xu*, Shounan An, Insoo Oh Netmarble AI Center Abstract This paper describes Netmarble’s submission to WMT21 Automatic Post-Editing (APE) Shared Task for the English-German language pair. First, we propose a Curriculum Training Strategy in training stages. Facebook Fair’s WMT19 news translation model was chosen to engage the large and powerful pre-trained neural networks. Then, we post-train the trans- lation model with different levels of data at each training stages. As the training stages go on, we make the system learn to solve multi- ple tasks by adding extra information at differ- ent training stages gradually. We also show a way to utilize the additional data in large volume for APE tasks. For further improve- ment, we apply Multi-Task Learning Strategy with the Dynamic Weight Average during the fine-tuning stage. To fine-tune the APE cor- pus with limited data, we add some related sub- tasks to learn a unified representation. Finally, for better performance, we leverage external translations as augmented machine translation (MT) during the post-training and fine-tuning. As experimental results show, our APE sys- tem significantly improves the translations of provided MT results by -2.848 and +3.74 on the development dataset in terms of TER and BLEU, respectively. It also demonstrates its ef- fectiveness on the test dataset with higher qual- ity than the development dataset. 1 Introduction Automatic Post-Editing (APE) aims to improve the quality of an existing Machine Translation (MT) system by learning from human-edited samples (Chatterjee et al., 2019, 2020). With the continu- ous performance improvements of Neural Machine Translation (NMT) systems along with deep learn- ing advancements, developing APE systems has faced a big challenge. Simple translation errors are hard to find in machine translation outputs, and ∗These authors equally contributed to this work. the remaining errors are still hard to solve. In re- cent years, transfer learning and data augmentation techniques have shown their efficiency when train- ing models on datasets with limited size (Devlin et al., 2019). Therefore, such approaches are also adopted in APE tasks (Lopes et al., 2019). Participants in WMT21 APE shared tasks are required to develop systems to automatically post- edit the translation outputs from an unknown MT system. In this year, the same data has been re- post-edited to improve the quality. As a result of performing statistics on the development set, the evaluation scores are 19.057 and 68.79 in terms of TER and BLEU, which are much higher than the scores of last year, 31.374 and 50.37, respec- tively. The central distribution of TER has shifted to the left compared to last year. We find that the section in the range of 5 to 10 has the most exam- ples, which indicates that over-correction problems should be considered during the APE tasks. In ad- dition, the dataset has been changed in terms of the …

Content truncated to 500 words