Word Level Robustness Enhancement: Fight Perturbation with Perturbation

State-of-the-art deep NLP models have achieved impressive improvements on many tasks. However, they are found to be vulnerable some perturbations. Before widely adopted, the fundamental issues of robustness need addressed. In this paper, we design a enhancement method defend against word substitution perturbation, whose basic idea is fight perturbation with perturbation. We find that: although well-trained not robust in setting presence adversarial samples, satisfy weak robustness. That means can handle most non-crafted perturbations well. Taking advantage property models, utilize resist crafted by attackers. Our contains two main stages. The first stage using randomized conform input data distribution. second eliminate instability prediction results and enhance guarantee. Experimental show that our significantly improve ability state-of-the-art attacks while maintaining performance original clean data.