A Matching-Integration-Verification Model for Multiple-Choice Reading Comprehension

Luxi Xing, Yue Hu, Yuqiang Xie, Chunhui Wang, Yong Hu

Abstract: Multiple-choice reading comprehension is a challenging task requiring a machine to select the correct answer from a candidate answers set. In this paper, we propose a model following a matching-integration-verification-prediction framework, which explicitly employs a verification module inspired by the human being and generates judgment of each option simultaneously according to the evidence information and the verified information. The verification module, which is responsible for recheck information from matching, can selectively combine matched information from the passage and option instead of transmitting them equally to prediction. Experimental results demonstrate that our proposed model achieves significant improvement on several multiple-choice reading comprehension benchmark datasets.