A three-part framework (asymptotic distillation, dynamic switch, rate-scheduled learning) fine-tunes BERT for NMT without catastrophic forgetting, yielding up to 2.9 BLEU gain on WMT14 En-De.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2019 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Towards Making the Most of BERT in Neural Machine Translation
A three-part framework (asymptotic distillation, dynamic switch, rate-scheduled learning) fine-tunes BERT for NMT without catastrophic forgetting, yielding up to 2.9 BLEU gain on WMT14 En-De.