Towards Improving Adversarial Training of NLP Models