Figure 1 Left results of different fine tuning baselines . Right Results of utilizing back translation on baselines

Common Response: We would like to thank all reviewers very much for the detailed feedback and valuable suggestions! We will follow the suggestions on writing and related works and revise accordingly. We agree with the reviewers' Net (AB-Net FB), the variant that trains AB-Net from scratch (AB-Net SC) and baselines trained with back-translation. We train all settings for 50 epochs and test on the validation set of IWSLT14 De-En. English Wikipedia data, which is a subset of the training corpus of BERT.