Tsinghua University Neural Machine Translation Systems for CCMT 2020

Gang Chen,Shuo Wang,Xuancheng Huang,Zhixing Tan,Maosong Sun,Yang Liu

DOI: https://doi.org/10.1007/978-981-33-6162-1_9

2020-01-01

Abstract:This paper describes the neural machine translation system of Tsinghua University for the bilingual translation task of CCMT 2020. We participated in the Chinese <-> English translation tasks. Our systems are based on Transformer architectures and we verified that deepening the encoder can achieve better results. All models are trained in a distributed way. We employed several data augmentation methods, including knowledge distillation, back-translation, and domain adaptation, which are all shown to be effective to improve translation quality. Distinguishing original text from translationese can lead to better results when performing domain adaptation. We found model ensemble and transductive ensemble learning can further improve the translation performance over the individual model. In both Chinese -> English and English -> Chinese translation tasks, our systems achieved the highest case-sensitive BLEU score among all submissions.