Universal Adversarial Triggers for Attacking and Analyzing NLP