知识图谱驱动的图卷积神经网络谣言检测模型  

Knowledge graph-driven graph neural network-based model for rumor detection

作　　者：徐凡[1] 李明昊 黄琪 鄢克雨 王明文[1] 周国栋[2] Fan XU;Minghao LI;Qi HUANG;Keyu YAN;Mingwen WANG;Guodong ZHOU(School of Computer Information Engineering,Jiangxi Normal University,Nanchang 330022,China;School of Computer Science&Technology,Soochow University,Suzhou 215006,China)

机构地区：[1]江西师范大学计算机信息工程学院,南昌330022 [2]苏州大学计算机科学与技术学院,苏州215006

出　　处：《中国科学：信息科学》2023年第4期663-681,共19页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:62162031,62076175,62066020,62266023);江西省杰出青年基金项目(批准号:20192ACBL21030)资助。

摘　　要：社交媒体谣言以极低的成本在互联网中被快速扩散,给社会带来显著的负面影响.传统的谣言检测模型主要考虑传播模式、写作风格、用户信用和世界知识等信息.然而,谣言的传播模式通常难以被捕捉,写作风格却容易被模仿,由元数据(如职业、家乡、学历、年龄等)构成的用户信息也容易被伪造.本文提出了一种新颖的知识驱动的图卷积神经网络谣言检测模型.该模型首先将社交媒体文本表示成一种语义–实体无向图结构,其中节点包含原社交媒体文本中的词语,利用世界知识库扩展的实体词语,以及利用语言知识库扩展的语义词语,边包含三类节点的6种有效组合.该语义–实体图可以有效地增强任意两种节点的共现性,从而丰富了原社交媒体文本的表示,从一定程度上缓解数据稀疏共现问题.语言知识利用了HowNet (义原和同义词)以及WordNet (上义词、下义词和同义词)分别对中英文社交媒体文本的主题词进行扩充.并成功地将语言知识和实体知识通过图卷积神经网络框架有效集成.在4个国际基准中英文谣言语料库上的实验结果和可视化分析表明了本文模型的有效性.Rumors can be propagated quickly across online social media at very low costs,resulting in a significant negative impact on society.The conventional rumor-detection models mainly consider information such as propagation patterns,writing style,user credit,and world knowledge.However,the propagation pattern of rumors is often difficult to capture,the writing style is easy to imitate,and the user information composed of metadata(e.g.,occupation,hometown,education,and age)is easy to forge.This paper presents a novel knowledge-driven graph convolutional network rumor-detection model.The model represents social media text as a semantic-entity undirected graph structure comprising edges containing six combinations of three types of nodes.The nodes,in turn,contain words from an original text,entity words extended by the world knowledge base,and semantic words extended by the language knowledge base.The semantic-entity graph can effectively enhance the co-occurrence between any two nodes to enrich the representation of the original social media text,alleviating the problem of sparse data co-occurrence.WordNet(hypernym,hyponym,and synonym)and HowNet(sememe and synonym)are adopted to extend the topic words of social media texts,and language and entity knowledge are successfully integrated through the framework of the graph convolution network.The experimental results based on four international benchmarks,Chinese and English databases,and visualization analyses show the effectiveness of our proposed model.

关 键 词：语言知识 世界知识 主题模型 图卷积神经网络 谣言检测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP391.1[自动化与计算机技术—控制科学与工程]