From Machine Reading Comprehension to Dialogue State Tracking: Bridging the Gap