Accurate Knowledge Distillation with n-best Reranking
We propose extending the Sequence-level Knowledge Distillation (Kim and Rush, 2016) with n-best reranking to consider not only the top-1 hypotheses but also the top n-best hypotheses of teacher models. Our approach leverages a diverse set of models, including publicly-available large pretrained models, to provide more accurate pseudo-labels for training student models. We validate our proposal on the WMT21 German-English translation task and demonstrate that our student model achieves comparable accuracy to a large translation model with 4.7 billion parameters from (Tran et al., 2021) while having two orders of magnitude fewer parameters.
READ FULL TEXT