The authors have presented a clear rebuttal and the reviewers have updated their reviews based on the rebuttal. There were still some aspects which the reviewers thought still needed to be discussed and shown in the paper. In discussion with the reviewers, we thought that given the empirical nature of the experiment design, interpretability and error analysis (as suggested by R3) are important. Less model architecture sensitivity analysis was compensated for by the authors showing that the proposed method can be applied to multiple tasks with reasonable performance. However, the magnitude of improvement wasn't obvious as the baseline comparison is not clear and completed.