A2T: Towards Improving Adversarial Training of NLP Models (EMNLP 2021 Findings) - View it on GitHub