EveMRC: Two-Stage Bidirectional Evidence Modeling for Multi-Choice Machine Reading Comprehension

Machine Reading Comprehension (MRC) requires a machine to answer questions after reading and comprehending the given documents. Multi-choice MRC is one of the most studied MRC tasks due to the convenience of evaluation and the diversity of question types. However, the interpretability of multi-choice MRC, especially in evidence extraction, remains underexplored because a correct answer may be derived by eliminating incorrect options rather than being supported by positive evidence. In this work, we propose a bidirectional evidence modeling framework, EveMRC, to enhance the explainability of Multi-choice MRC systems. Compared to previous works, our framework exclusively addresses the problem of bidirectional evidence selection, which not only selects positive evidence for the right answer but also selects negative evidence for wrong answers. The bidirectional evidence can also facilitate model decisions by incorporating it into a competition process. To avoid the high annotation cost of bidirectional evidence, our framework utilizes a novel weakly-supervised pipeline to train the evidence selector. Experimental results on four multi-choice MRC datasets demonstrate the effectiveness of our framework, which not only enhances the explainability of MRC systems but also improves their overall performance.