TextAT: Adversarial Training for Natural Language Understanding with Token-Level Perturbation

Adversarial training is effective in improving the robustness of neural networks. In NLP, languages are discrete in nature, separate tokens possess discrete semantics. Therefore, to incorporate adversarial training in sequence-level tasks, we introduce a novel training strategy: Text Adversarial Training with token-level perturbation. We fist craft perturbations that are initialized using a fine-grained token-level accumulated perturbations. Then we constrain these perturbations considering that inputs are separate tokens, rather than constraining them under a naive normalization ball. We validate the effectiveness of such normalization method using large-scale Transformer-based language models. Experiments on GLUE benchmark and NER task show that our adversarial training strategy improves the performances on various tasks including text classification and sequence labeling.