基于推文传播模式与跨模态特征的网络谣言检测研究  

Research on Online Rumor Detection Based on Tweet Propagation Patterns and Cross-Modal Features

作　　者：彭竞杰 顾益军[1] 张岚泽 Peng Jingjie;Gu Yijun;Zhang Lanze(Department of Information and Network Security,People's Public Security University of China,Beijing 100038)

机构地区：[1]中国人民公安大学信息网络安全学院,北京100038

出　　处：《情报杂志》2024年第8期124-134,58,共12页Journal of Intelligence

基　　金：中央高校建设世界一流大学(学科)和特色发展引导专项“中国人民公安大学网络空间安全执法技术双一流创新研究项目”(编号:2023SYL07)研究成果。

摘　　要：[研究目的]为了有效治理网络谣言,减少网络谣言对社会稳定带来的威胁,提出充分整合帖子的多模态信息和传播模式信息对谣言进行精准识别。[研究方法]提出融合推文传播模式信息与跨模态特征的网络谣言检测模型(PPCMRD)。在推文传播特征挖掘方面,首先通过推断潜在连接补全推文传播图,接着采用双向标签图注意力模块编码推文的多个传播模式,然后通过传播模式信息融合模块捕获模式特征间的互补信息,得到帖子的传播特征;在整合多模态特征方面,该模型将帖子的文本、图像和推文传播特征集成在一起,采用跨模态共同注意力机制捕捉不同模态信息间的互补关系,得到帖子的最终嵌入表示,判断是否是谣言。[研究结论]在两个公开数据集上的实验结果表明,PPCMRD模型能够有效地检测谣言,并优于当前的基线模型。[Research purpose]To effectively manage online rumors and reduce the threat of online rumors to social stability,we propose to fully integrate the multimodal information of tweets and the propagation pattern information to accurately identify rumors.[Research method]We propose a rumor detection model(PPCMRD)that integrates tweet propagation pattern information and cross-modal features.In terms of tweet propagation feature mining,the first step is to complement the tweet propagation graph by inferring potential connections,followed by encoding multiple propagation patterns of tweets using the bidirectional signed graph attention module,and then capturing the complementary information between pattern features through the propagation pattern information fusion module to obtain the propagation features of the tweet.In terms of integrating the multimodal features,this model integrates the text,image,and tweet propagation features of the tweet,and employs the cross-modal co-attention mechanism to capture the complementary relationship between different modal information and get the final embedding representation of the tweet to determine whether it is a rumor or not.[Research conclusion]The experimental results on two public datasets demonstrate that the proposed approach could effectively detect rumors and outperforms the current baseline models.

关 键 词：网络谣言 谣言检测 网络谣言检测模型 推文传播模式信息 跨模态特征融合 

分 类 号：G252[文化科学—图书馆学]