machine reading comprehension (MRC) is an important area of conversation
agents and draws a lot of attention. However, there is a notable limitation to
current MRC benchmarks: The labeled answers are mostly either spans extracted
from the target corpus or the choices of the given candi