MATE-KD: Masked Adversarial TExt, a Companion to Knowledge Distillation

TL;DR

MATE-KD introduces a novel adversarial training method that enhances knowledge distillation for NLP models by perturbing text to improve student model performance, outperforming existing baselines on GLUE.

Contribution

The paper proposes MATE-KD, a new adversarial training algorithm that improves knowledge distillation by using a masked language model to generate perturbed training samples.

Findings

MATE-KD outperforms competitive adversarial and data augmentation baselines.

A 6-layer RoBERTa model with MATE-KD surpasses BERT-Large on GLUE.

The method effectively enhances student model performance through text perturbation.

Abstract

The advent of large pre-trained language models has given rise to rapid progress in the field of Natural Language Processing (NLP). While the performance of these models on standard benchmarks has scaled with size, compression techniques such as knowledge distillation have been key in making them practical. We present, MATE-KD, a novel text-based adversarial training algorithm which improves the performance of knowledge distillation. MATE-KD first trains a masked language model based generator to perturb text by maximizing the divergence between teacher and student logits. Then using knowledge distillation a student is trained on both the original and the perturbed training samples. We evaluate our algorithm, using BERT-based models, on the GLUE benchmark and demonstrate that MATE-KD outperforms competitive adversarial learning and data augmentation baselines. On the GLUE test set our 6 layer RoBERTa based model outperforms BERT-Large.