Beat the AI: Investigating Adversarial Human Annotation for Reading Comprehension