EveMRC: A Two-stage Evidence Modeling For Multi-choice Machine Reading Comprehension

Anonymous

Abstract: Many impressive works have been proposed to improve the performance of Machine Reading Comprehension (MRC) systems in recent years. However, it is still difficult to interpret the predictions of existing MRC models, which makes the predictions unconvincing.In this work, we propose a two-stage explainable framework for multi-choice MRC to model not only the correlation between answers and evidence, but also the competition among evidence. In stage 1, we select evidence sentences for both the right answer and wrong answers using the semi-supervised evidence selector. In stage 2, we employ an evidence discriminator to compare among the competitive evidence set and make final judgments. Moreover, we propose an evidence-enabled data augmentation method. Experiments on four multi-choice MRC datasets show that: stage 1 provides strong explainability for MRC systems and stage 2 improves both the performance and robustness of MRC systems meanwhile.