基于时间步局部动态交互的多任务谣言检测方法  

Multi-Task Rumor Detection Method Based on Time-Step Dynamic Interaction

作　　者：杨广浩 万书振 董方敏[1,2] 王梦园[1,2] YANG Guanghao;WAN Shuzhen;DONG Fangmin;WANG Mengyuan(Hubei Key Laboratory of Intelligent Vision Based Monitoring for Hydroelectric Engineering,China Three Gorges University,Yichang,Hubei 443002,China;College of Computer and Information Technology,China Three Gorges University,Yichang,Hubei 443002,China)

机构地区：[1]三峡大学湖北省水电工程智能视觉监测重点实验室,湖北宜昌443002 [2]三峡大学计算机与信息学院,湖北宜昌443002

出　　处：《计算机工程与应用》2025年第6期183-191,共9页Computer Engineering and Applications

基　　金：国家自然科学基金新疆联合基金重点项目(U1703261,U61871258);三峡库区生态环境教育部工程研究中心开放基金项目(KF2023-11)。

摘　　要：谣言检测旨鉴别社交媒体中未经官方证实或人为捏造的信息,而当今社交网络中隐含着一种难以发掘的动态关系模式,它随时间推移和不同帖子间的动态交互而变化。针对现有方法对谣言传播事件中隐含的动态特征和关联信息考虑不充分的问题,提出一种基于时间步局部动态交互的多任务谣言检测方法,能捕获谣言传播事件中隐含的动态关联信息;并设计了一种高效的多任务交互方式,以时间步为基本共享单元,将学习到的局部特征进行共享,极大提升了共享效率,从而形成局部动态交互,整体多任务共享的检测框架。最后利用注意力机制筛选不同任务、不同结构特征中对谣言检测更有利的信息,以提升检测效果。在PHEME和WEIBO数据集上进行了实验,结果表明该方法具有较先进的性能。Rumour detection aims to identify information on social media that has not been officially verified or artificially created.Today’s social networks contain a pattern of dynamic relationships that are difficult to find,changing over time and dynamic interactions between different posts.This paper proposes a multi-task rumour detection method based on time step local dynamic interaction,which can capture the hidden dynamic correlation information in rumour spreading events.An efficient multi-task interaction method is designed,in which time step is used as the basic sharing unit,the local features learned are shared,and the sharing efficiency is greatly improved,so that the detection framework of local dynamic interaction and whole multi-task sharing is formed.Finally,the attention mechanism is used to filter the information that is more favorable for rumour detection in different tasks and different structural characteristics to improve the detection effect.Experiments on PHEME and WEIBO datasets show that this method has advanced performance.

关 键 词：谣言检测 时间步局部动态交互 传播结构特征 多任务共享 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]