Evaluating Paraphrastic Robustness in Textual Entailment Models