Impacts of Multi-task Adversarial Training on Machine Reading Comprehension Model

This paper investigates multi-task adversarial training to enhance the generalization and robustness of machine reading comprehension (MRC) models. Moreover, experiments are carried out to quantitatively analyze the interference to the model by different adversarial sample ratio, which aims to further improve the adversarial training strategy to take maximum advantage of adversarial samples. The experimental results show that the MRC model trained by the proposed multi-task adversarial training strategy has better robustness than several recent state-of-the-art (SOTA) models when facing adversarial attacks.