Lang Qi, Liu Xiaodong, Jia Wenjuan
IEEE Trans Neural Netw Learn Syst. 2023 Dec;34(12):10904-10918. doi: 10.1109/TNNLS.2022.3171677. Epub 2023 Nov 30.
Open-domain question answering (QA) tasks require a model to retrieve inference chains associated with the answer from massive documents. The core of a QA model is the information filtering ability and reasoning ability. This article proposes a semantic knowledge reasoning graph model based on the multidimensional axiomatic fuzzy set (AFS), which can generate the knowledge graph (KG) and build reasoning paths for reading comprehension tasks through unsupervised learning. Moreover, taking advantage of the interpretable AFS framework enables the proposed model to have the ability to learn and analyze the semantic relationships between candidate documents. Meanwhile, the utilization of the multidimensional AFS acquires semantic descriptions of candidate documents more concise and flexible. The similarity degree between paragraphs is calculated according to the AFS description to generate the graph. Interpretable chains of reasoning provided by the AFS knowledge graph (AFS Graph) will serve as the basis for the answer prediction. Compared with the previous methods, the AFS Graph model presented in this article improves interpretability and reasoning ability. Experimental results show that the proposed model can achieve the state-of-the-art performance on datasets of HotpotQA, SQuAD, and Natural Questions Open.
开放域问答（QA）任务要求模型从海量文档中检索与答案相关的推理链。QA模型的核心是信息过滤能力和推理能力。本文提出了一种基于多维公理模糊集（AFS）的语义知识推理图模型，该模型可以通过无监督学习生成知识图（KG）并为阅读理解任务构建推理路径。此外，利用可解释的AFS框架使所提出的模型具有学习和分析候选文档之间语义关系的能力。同时，多维AFS的使用使得候选文档的语义描述更加简洁和灵活。根据AFS描述计算段落之间的相似度以生成图。AFS知识图（AFS Graph）提供的可解释推理链将作为答案预测的基础。与先前的方法相比，本文提出的AFS Graph模型提高了可解释性和推理能力。实验结果表明，所提出的模型在HotpotQA、SQuAD和Natural Questions Open数据集上可以实现最优性能。
