背景知识增强的多特征融合谣言检测方法  

Multi-feature Fusion for Rumor Detection with Background Knowledge Enhancement

作　　者：林兴澎 李家印 徐瑞阳 许力 LIN Xingpeng;LI Jiayin;XU Ruiyang;XU Li(College of Computer and Cyber Security,Fujian Normal University,Fuzhou 350117,China;Fujian Provincial Key Laboratory of Network Security and Cryptology,Fuzhou 350117,China)

机构地区：[1]福建师范大学计算机与网络空间安全学院,福州350117 [2]福建省网络安全与密码技术重点实验室,福州350117

出　　处：《小型微型计算机系统》2025年第3期559-570,共12页Journal of Chinese Computer Systems

基　　金：国家自然科学基金NSFC海峡联合基金项目(U1905211)资助;福建省科技项目(2022G02003,2023L3007)资助.

摘　　要：在线社交媒体的普及为人们通信带来便利,但也为谣言滋生创造条件,设计高效的谣言检测方法能保护人民财产和维持社会稳定.已有方法主要集中在利用谣言传播中的丰富信息来检测谣言,这些方法在长期谣言检测具有优越性能,但应对早期谣言检测的效果不佳.针对这些方法无法在谣言传播早期获得丰富信息的问题,本文提出了一种背景知识增强的多特征融合谣言检测方法来提高早期谣言检测性能.首先,从知识图谱和维基百科中挖掘谣言背景知识并建立知识关联图来补充源推文的语义信息;其次,为了解决现有方法难以学习具有不同差异性噪声的谣言传播表示的问题,本文设计了一种基于加性注意力和点积注意力的图神经网络结构对谣言进行插值学习;最后,将知识关联图、谣言传播-扩散图以及社交图的表示进行结合,构建出具有多通道输入的谣言检测器架构,从而实现早期谣言的精准分类.实验结果表明,本文方法在3个公开数据集上的准确率分别达到了87.3、90.4%和87.0%,与其它对比方法相比,具有更高的早期谣言检测准确率和长期谣言检测准确率.The popularity of online social media brings convenience to people′s communication,but it also creates conditions for rumors to breed.Designing efficient rumor detection can protect property and maintain social stability.It has been mainly focused on using the rich information in the spread of rumor to detect rumors,which has superior performance in rumor detection,but the effect of dealing with early rumor detection is not good.To solve these problems,a multi-feature fusion for rumor detection with background knowledge enhancement is proposed to improve the performance of early rumor detection.Firstly,the rumor background knowledge is mined from the knowledge graph and the knowledge association graph is established to supplement the semantic information of the source tweet;Secondly,in order to solve the problem that it is difficult to learn the representation of rumor propagation with different noise,this paper designs a graph neural network structure based on additive attention and dot product attention to learn rumors;Finally,the knowledge association graph,rumor propagation-diffusion graph and social graph are combined to construct a rumor detector architecture with multi-channel input,so as to achieve the accurate classification of early rumors.Experimental results show that the accuracy of this work on three public datasets is 87.3%,90.4%,and 87.0%,respectively.Compared to other comparison methods,it has higher accuracy in early rumor detection and long-term rumor detection.

关 键 词：社交媒体 谣言检测 图神经网络 知识图谱 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]