| 摘要: |
| 随着谣言在网络上不断扩散和传播,其危害会越来越大。在谣言尚未传播的最早期,利用源发布账号的用户信息和文本信息对其进行识别并遏止具有十分重要的意义。当前的检测方法局限于自然语言处理技术,注重从文本中提取信息来识别谣言,缺乏对用户信息的深度挖掘和有效结合,导致模型的检测性能低。为此,本文提出了一种谣言早期检测的新途径RT-NIG,通过在交叉分布的邻域信息图中重构二元组信息来识别谣言。首先针对谣言传播最初阶段缺乏传播信息,无法形成图结构数据的情景,利用对象的潜在相关性构造虚拟邻域图,解决了数据不确定性以及不完备性等问题;通过图神经网络捕获邻域图中潜在的对象关系,关注用户之间潜在的可信度关系以及文本之间的情感极性关系,在两个邻域信息图中交叉传递用户信息和语义信息,分别重构了用户信息和语义信息;最后通过加权集成的方式,重新构造“用户-推文”二元组信息,有效地结合了这两种信息,并用于下游的谣言分类任务。分别在中文Weibo和英文PHEME两个真实数据集上进行了实验验证,本文方法在准确率、精确率、召回率、F1值等指标上优于多种先进的早期检测方法,在两个数据集上准确率分别比最优的对比方法提升了5%和8%;并且通过消融研究以及超参数分析,进一步证明了用户信息在早期检测中具有的重要作用以及二元组信息重构方式的有效性。针对无传播信息可用的场景,RT-NIG也为其他一些早期检测问题提供了新的解决方法,例如假新闻、网络暴力,误导消息等问题。 |
| 关键词: 谣言早期检测 RT-NIG 邻域信息图 重构二元组 图神经网络 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2024.11.10 |
| 投稿时间：2022-12-08修订日期：2023-05-10 |
| 基金项目:本课题得到嵩山实验室项目(No.221100210700-2)和河南省重大科技专项(No.221100210100)的资助。 |
|
| RT-NIG: Reconstructing Two-tuples in Neighborhood Information Graphs for Early Rumor Detection |
| WANG Biao,WEI Hongquan,LIU Shuxin,WANG Kai,JIANG Haocong,LI Ran |
| PLA Strategic Support Force Information Engineering University, Zhengzhou 450001, China;PLA Strategic Support Force Information Engineering University, Zhengzhou 450001, China;National Digital Switching System Engineering and Technological R&D Center, Zhengzhou 450002, China |
| Abstract: |
| As rumors continue to diffuse and spread on the network, their influence will become more and more serious. In the earliest stage that rumors have not yet spread, it is of great significance to use the user information and text information of the source account to identify and suppress them. The current detection methods are limited to natural language processing technology and focus on extracting information from the text to identify rumors. The lack of in-depth mining and an effective combination of user information leads to low detection performance of the model. In this paper, we propose a new approach for early rumor detection, RT-NIG, which can identify rumors by reconstructing two-tuples information in the cross-distributed neighborhood information graphs. Firstly, in response to the situation where there is a lack of dissemination information in the initial stage of rumor propagation and graph-structured data cannot be formed, a virtual neighborhood graph is constructed using the potential correlation of objects to solve the problems of data uncertainty and incompleteness. Then the potential object relationships in the neighborhood graph are captured by graph neural network. The user information and semantic information are transferred in two neighborhood information graphs and reconstructed, paying attention to the potential credibility relationship between users and the emotional polarity relationship between texts. Finally, using weighted integration, the "User-Tweet" two-tuple information is reconstructed, effectively combining the two kinds of information, and is used for the downstream rumor classification task. Experiments are conducted on two real datasets, Chinese Weibo and English PHEME. The proposed method outperformed various advanced early detection methods in accuracy, accuracy, recall, F1 value, and other indicators. The accuracy was improved by 5% and 8% compared to the optimal comparison method on the two datasets. Ablation research and super parameter analysis further prove the important role of user information in early detection and the effectiveness of two-tuple information reconstruction. RT-NIG also provides new solutions for other early detection problems, such as fake news, online violence, misleading information,etc., in scenarios where no disseminated structural information is available. |
| Key words: early rumor detection RT-NIG neighborhood information graphs reconstruct two-tuples GNN |
