Interpretable modular knowledge reasoning for machine reading comprehension

Machine reading comprehension (MRC) is a fundamental task of evaluating the natural language understanding ability of model, which requires complicated reasoning about the knowledge involved in the context as well as world knowledge. However, most existing approaches ignore the complicated reasoning process and solve it with a one-step "black box" model and massive data augmentation. Therefore, in this paper, we propose a modular knowledge reasoning approach based on neural network modules that explicitly model each reasoning process step. Five reasoning modules are designed and learned in an end-to-end manner, which leads to a more interpretable model. Experiments using the reasoning over paragraph effects in situations (ROPES) dataset, a challenging dataset that requires reasoning over paragraph effects in a situation, demonstrate the effectiveness and explainability of our proposed approach. Moreover, the transfer of our reasoning modules to the WinoGrande dataset under the zero-shot setting achieved competitive results compared with the data augmented model, proving the generalization capability.