Residual Connection-Based Multi-step Reasoning via Commonsense Knowledge for Multiple Choice Machine Reading Comprehension

Abstract

Generally, the candidate options for multiple choice machine reading comprehension (MRC) are not explicitly present in the document and need to be inferred from text or even from the world’s knowledge. Previous work endeavored to improve performance with the aid of commonsense knowledge or using multi-step reasoning strategy. However, there is no model adopt multi-step reasoning with external commonsense knowledge information to solve multiple choice MRC, and two shortcomings still remain unsolved, i.e., external knowledge may involve undesirable noise and only the latest reasoning step makes contribution to the next reasoning. To address the above issues, we propose a multi-step reasoning neural network based on the strong Co-Matching model with the aid of commonsense knowledge. Firstly, we present a sentence-level knowledge interaction (SKI) module to integrate commonsense knowledge with corresponding sentence rather than the whole MRC instance. Secondly, we present a residual connection-based multi-step reasoning (RCMR) answer module, which makes the next reasoning depending on the integration of several early reasoning steps rather than only the latest reasoning step. The comparative experimental results on MCScript show that our single model achieves a promising result comparable to SOTA single model with extra samples and specifically achieves the best result for commonsense type questions.