AMAMP: A Two-Phase Adaptive Multi-hop Attention Message Passing Mechanism for Logical Reasoning Machine Reading Comprehension

Current machine reading comprehension datasets that involve logical reasoning have gained significance in Natural Language Processing. The challenge with these datasets lies in requiring models to comprehend and effectively reason through logical questions, where logical sentences are distantly positioned and lack direct connections. To tackle this challenge, we employ graph diffusion and adaptive mechanisms, aiming to enhance the extraction of information among distant nodes while efficiently mitigating noise during edge expansion and inter-node information propagation. Leveraging advanced Graph Neural Networks, our models - MAMP, LoFiMAMP, and AMAMP - demonstrate effective inference for logical questions by facilitating information exchange among long-distance nodes. Experimental results on the ReClor and LogiQA datasets demonstrate high accuracy, 53.1% and 34.6%, that surpasses all baseline models.