融合多层级特征表示的多领域谣言早期检测方法  

A Multi-Domain Rumor Early Detection Method Fusing Multi-Level Feature Representations

作　　者：黄涛 肖玉芝[1,2,3] 向洁萍 金胜 霍宣蓉 Huang Tao;Xiao Yuzhi;Xiang Jieping;Jin Sheng;Huo Xuanrong(Qinghai Normal University,Xining 810016;Tibetan Information Processing and Machine Translation Key Laboratory of Qinghai Province,Xining 810008;Key Laboratory of Tibetan Information Processing Ministry of Education,Qinghai Normal Universty,Xining 810008)

机构地区：[1]青海师范大学计算机学院,西宁810016 [2]青海省藏文信息处理与机器翻译重点实验室,西宁810008 [3]青海师范大学藏文信息处理教育部重点实验室,西宁810008

出　　处：《情报杂志》2025年第4期127-135,共9页Journal of Intelligence

基　　金：国家重点研发计划项目“数据融合规范及原型”(编号:314);国家重点实验室自主项目“社交媒体失真信息识别及溯源关键技术研究”(编号:2024-SKL-005)研究成果。

摘　　要：[研究目的]网络谣言的治理是当前社会广泛关注的问题,提高网络谣言在传播早期的识别效率,能更好的阻止谣言信息的传播并维护社会的和谐稳定。[研究方法]提出一种多领域话题下的早期谣言检测方法。通过协同注意力机制融合文本的词汇、短语和句子级特征,构建多层级特征增强的单元门模块以挖掘谣言深层信息。利用该模块构建领域感知特征抽取器,捕获谣言文本的领域特征及偏差,形成多领域与多层级的谣言特征表示,判断是否为谣言。[研究结果/结论]在涵盖9个不同领域的公开数据集上的实验结果表明,该模型的准确率、F1值和AUC值分别达到了92.85%、93.11%和96.96%,能够有效的对多领域谣言进行早期检测。[Research purpose]The management of network rumors is a widely concerned issue in the current society.Improving the efficiency of identifying network rumors at the early stage of their dissemination can better stop the dissemination of rumor information and maintain the harmony and stability of the society.[Research method]A method for early rumor detection across multiple domains is proposed.Initially,a multi-level feature-enhanced unit gate module is constructed by integrating the word,phrase,and sentence-level features of the text through a cooperative attention mechanism,aiming to delve into the deep information of rumors.Subsequently,this module is utilized to build a domain-aware feature extractor,capturing the domain-specific features and biases of rumor texts.This results in a multi-domain and multi-level representation of rumor features,which can be used to determine whether a piece of information is a rumor.[Research result/conclusion]Experimental results on publicly available datasets covering nine different domains show that the accuracy,F1 value and AUC value of this paper's model reach 92.85%,93.11%and 96.96%,respectively,and are able to effectively perform eary detection of multi-domain rumors.

关 键 词：网络谣言 谣言识别 早期谣言检测 多领域话题 特征增强 领域感知 

分 类 号：TP39[自动化与计算机技术—计算机应用技术] G250[自动化与计算机技术—计算机科学与技术]