LoCSGN: Logic-Contrast Semantic Graph Network for Machine Reading Comprehension

Abstract

AbstractLogical reasoning plays a very important role in Natural Language Understanding. In recent years, research on logical reasoning has shown a booming trend. Previous works either construct graphs to perform implicit reasoning or augment labeled training data with consistent symbol logical rules. In this paper, we combine the advantages of both graph based implicit reasoning and symbolic logic based consistent transformation in a deep learning framework. In order to make full use of the semantic and logical structures in text, we exploit Abstract Meaning Representation (AMR) to help logical reasoning, which can explicitly provide core semantic knowledge and logical structures. Based on AMR graph extracted, we design two tasks: 1) Construct joint graph and strengthen the interaction between context and option subgraph to predict right choice. 2) Leverage symbolic rules to construct logical consistent and inconsistent graphs to let model identify and differentiate logical structures in different graphs by contrastive learning. Experiments are conducted on two logical reasoning datasets: Reclor and LogiQA. And our method has a significant improvement over the baselines and most previous methods.KeywordsMachine reading comprehensionLogic reasoningContrastive learning