Universal Adversarial Perturbation for Text Classification