基于贴文级特征融合的社交网络谣言检测方法  

SOCIAL NETWORKS RUMOR DETECTION APPROACH BASED ON POST-LEVEL FEATURE FUSION

作　　者：余潇龙 郭天成 陈阳[1,2] 王新 Yu Xiaolong;Guo Tiancheng;Chen Yang;Wang Xin(School of Computer Science,Fudan University,Shanghai 201203,China;Shanghai Key Laboratory of Intelligent Information Processing,Shanghai 201203,China)

机构地区：[1]复旦大学计算机科学技术学院,上海201203 [2]上海市智能信息处理重点实验室,上海201203

出　　处：《计算机应用与软件》2024年第8期189-195,共7页Computer Applications and Software

基　　金：上海市自然科学基金项目(16ZR1402200)。

摘　　要：现有的谣言检测方法,对贴文语义和贴文发布者信息、贴文传播状态等社交网络属性之间的关联利用不足,导致检测性能受限。针对这一不足,提出一种基于贴文级特征融合的谣言检测模型PF-HAN。模型通过带注意力机制的双向长短期记忆网络(Bi-LSTM)生成各条贴文的语义表示,并与对应贴文的社交网络特征进行拼接以保留两者的对应关系。融合得到的贴文综合表示以序列形式输入分层注意力网络提取时序特征,生成最终的事件表示用于谣言判别。实验结果表明,模型在新浪微博和Twitter数据集中进行谣言检测任务时F1值达到0.956和0.740,且能以高准确率完成谣言早期检测任务。The existing rumor detection methods largely neglect the correlation between post semantics,post publishers and post propagation status,which lead to low detection rates.To solve this problem,this paper proposes a rumor detection approach PF-HAN based on post-level feature fusion.The model used a Bi-LSTM with attention mechanism to generate the semantic representation of each post,and spliced it with the social network features of the corresponding post to preserve the correspondence between them.The integrated representation of the posts obtained by the fusion was input into the hierarchical attention network in the form of sequence to extract the temporal features and generate the final event representation for rumor discrimination.Experimental results over Sina Weibo and Twitter show that the F1 value of the model reaches 0.956 and 0.740 when the model performs the rumor detection task and it can complete the early rumor detection task with high accuracy.

关 键 词：深度学习 自然语言处理 谣言检测 特征融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]