Robust Domain Adaptation for Machine Reading Comprehension

Most domain adaptation methods for machine reading comprehension (MRC) use a pre-trained question-answer (QA) construction model to generate pseudo QA pairs MRC transfer. Such process will inevitably introduce mismatched (i.e., Noisy Correspondence) due i) the unavailable in target documents, and ii) shift during applying domain. Undoubtedly, noisy correspondence degenerate performance of MRC, which however is neglected by existing works. To solve such an untouched problem, we propose construct additionally using dialogue related as well new method MRC. Specifically, Robust Domain Adaptation Machine Reading Comprehension (RMRC) consists answer extractor (AE), question selector (QS), model. RMRC filters out irrelevant answers estimating correlation document via AE, extracts questions fusing candidate multiple rounds chats QS. With extracted pairs, fine-tuned provides feedback optimize QS through novel reinforced self-training method. Thanks optimization QS, our greatly alleviate problem caused shift. best knowledge, this could be first study reveal influence models show feasible solution achieve robustness against pairs. Extensive experiments on three datasets demonstrate effectiveness