Self-distillation distills the deeper sub-networks to the shallower sub-networks without using an extra teacher model, and has been proven effective in improving the performance of a series of computer vision tasks. In this paper, we study the representation-based self-distillation methods for Neural Machine Translation (NMT) considering the efficiency issue with a large vocabulary. We present a rank-order augmented Pearson correlation loss and an iterative distillation method to prevent the discrepancy of predictions between the student and a stronger teacher from disturbing the training. To prevent the teacher from misleading the student’s learning, we utilize a warm-up strategy and present a gradient adaption method to scale down or zero the Knowledge Distillation (KD) gradients which are opposite to the translation. Experiments show that our method can lead to significant improvements over the strong Transformer baseline on low/middle/high-resource tasks, obtaining comparable performance to previous MT KD studies without pre-training a teacher. Deeper Transformer experiments show that our method can lead to comparable or better performance with fewer layers.

Sequence-to-sequence (seq2seq) models achieve comparable or better grammatical error correction performance compared to sequence-to-edit (seq2edit) models. Seq2edit models normally iteratively refine the correction result, while seq2seq models decode only once without aware of subsequent tokens. Iteratively refining the correction results of seq2seq models via Multi-Pass Decoding (MPD) may lead to better performance. However, MPD increases the inference costs. Deleting or replacing corrections in previous rounds may lose useful information in the source input. We present an early-stop mechanism to alleviate the efficiency issue. To address the source information loss issue, we propose to merge the source input with the previous round correction result into one sequence. Experiments on the CoNLL-14 test set and BEA-19 test set show that our approach can lead to consistent and significant improvements over strong BART and T5 baselines (+1.80, +1.35, and +2.02 F0.5 for BART 12-2, large and T5 large respectively on CoNLL-14 and +2.99, +1.82, and +2.79 correspondingly on BEA-19), obtaining F0.5 scores of 68.41 and 75.36 on CoNLL-14 and BEA-19 respectively.

Paraphrasing, i.e., restating the same meaning in different ways, is an important data augmentation approach for natural language processing (NLP). Zhang et al. (2019b) propose to extract sentence-level paraphrases from multiple Chinese translations of the same source texts, and construct the PKU Paraphrase Bank of 0.5M sentence pairs. However, despite being the largest Chinese parabank to date, the size of PKU parabank is limited by the availability of one-to-many sentence translation data, and cannot well support the training of large Chinese paraphrasers. In this paper, we relieve the restriction with one-to-many sentence translation data, and construct ParaZh-22M, a larger Chinese parabank that is composed of 22M sentence pairs, based on one-to-one bilingual sentence translation data and machine translation (MT). In our data augmentation experiments, we show that paraphrasing based on ParaZh-22M can bring about consistent and significant improvements over several strong baselines on a wide range of Chinese NLP tasks, including a number of Chinese natural language understanding benchmarks (CLUE) and low-resource machine translation.