On Adversarial Examples for Text Classification by Perturbing Latent Representations