基于知识图谱与门控机制的专家再学习推理问答方法  

Expert Relearning Reasoning Question Answering Method Based on Knowledge Graph and Gating Mechanism

作　　者：房晓 王红斌[1,2,3] FANG Xiao;WANG Hongbin(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650500,Yunnan,China;Yunnan Provincial Key Laboratory of Artificial Intelligence,Kunming University of Science and Technology,Kunming 650500,Yunnan,China;Yunnan Key Laboratory of Computer Technology Application,Kunming University of Science and Technology,Kunming 650500,Yunnan,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650500 [2]昆明理工大学云南省人工智能重点实验室,云南昆明650500 [3]昆明理工大学云南省计算机技术应用重点实验室,云南昆明650500

出　　处：《应用科学学报》2025年第2期288-300,共13页Journal of Applied Sciences

基　　金：国家自然科学基金(No.61966020);云南省基础研究计划面上项目(No.CB22052C143A)资助。

摘　　要：现有使用预训练语言模型和知识图谱的图神经网络问答的方法主要集中于构建知识图谱子图及推理过程的研究,这类方法忽略了问题上下文与知识图谱的语义差异,不能深层次挖掘文本表示形式与知识图谱表示形式的语义特征,且缺失两种表示形式的知识源对答案预测贡献度不同的综合考虑。针对上述问题,本文提出了一种基于知识图谱与门控机制的专家再学习推理问答方法。该方法将问题上下文表示及推理后的知识图谱表示进行拼接融合,并将融合后的表示向量随机分配至专家网络中,再次学习问题上下文与知识图谱所关联的实体语义特征来挖掘深层隐含知识,并结合门控机制对问题上下文及推理后的知识图谱表示精准打分,通过动态调整两种表示形式的知识源对答案预测的贡献,提升答案预测精度。在CommonsenseQA数据集和OpenBookQA数据集上进行了实验,实验结果表明所提方法的准确率比QA-GNN方法分别提高了2.08%和1.23%。Existing graph neural network(GNN)-based question answering(QA)methods using pre-trained language models and knowledge graphs mainly focus on building knowledge graph subgraphs and reasoning processes.However,such methods ignore the semantic differences between question context and knowledge graphs,limiting their ability to deeply mine text representations.Moreover,they fail to comprehensively consider the varying contributions of these two representations to answer prediction.To address these challenges,this paper proposes an expert relearning reasoning QA method based on knowledge graphs and a gating mechanism.This method splices and fuses the question context representation with the inferred knowledge graph representation,and randomly assigns the fused representation vector to the expert network to relearn the entity semantic features associated with the question context and knowledge graph.By mining deeper hidden knowledge and incorporating the gating mechanism,the model accurately scores the question context and the inferred knowledge graph representation,dynamically adjusting their contribution to the answer prediction,and improving prediction accuracy.The proposed method was tested on the CommonsenseQA dataset and OpenBookQA dataset,achieving accuracy improvements of 2.08%and 1.23%over the QA-GNN method,respectively.

关 键 词：推理问答 知识图谱 图神经网络 门控机制 专家网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]