Repairing Adversarial Texts through Perturbation