一种基于深度学习的谣言检测方法  

A Rumor Detection Method Based on Deep Learning

作　　者：姜敏敏[1] 班浩 赵力[2] JIANG Minmin;BAN Hao;ZHAO Li(School of Network and Communication,Nanjing Vocational College of Information Technology,Nanjing Jiangsu 210023,China;School of Information Science and Engineering,Southeast University,Nanjing Jiangsu 210096,China)

机构地区：[1]南京信息职业技术学院网络与通信学院,江苏南京210023 [2]东南大学信息科学与工程学院,江苏南京210096

出　　处：《电子器件》2022年第6期1429-1433,共5页Chinese Journal of Electron Devices

基　　金：江苏省“青蓝工程”优秀青年骨干教师培养项目。

摘　　要：为了更好地学习网络谣言传播过程中的特征变化,提出了一种基于多跳的多模态融合的网络谣言检测方法。该方法采用faster RCNN提取视觉特征,通过GRU提取词特征,通过BERT提取句子特征,在提取词句基本特征后,利用RGCN实现图中不同节点间的信息传递。提取多模态特征后利用多跳注意力机制实现谣言检测。该方法可以较好解决诸如否定、歧义和长距离依赖等复杂问题,可以在更短路径上捕获远程依赖。通过与其他谣言检测方法的对比实验,验证了该方法在谣言检测,甚至早期谣言检测领域应用的有效性。In order to better learn the characteristic changes in the process of network rumor propagation,a network rumor detection method based on multi hop multimodal fusion is proposed.Faster RCNN is used to extract visual features,GRU is used to extract word features,and BERT is used to extract sentence features.After extracting basic features of words and sentences,RGCN is used to realize information transmission between different nodes in the graph.After extracting multimodal features,multi hop attention mechanism is used to detect rumors.This method can solve complex problems such as negation,ambiguity and long-distance dependency,and can capture remote dependency in a shorter path.By comparing with other rumor detection methods,the effectiveness of this method in the field of rumor detection and even early rumor detection is verified.

关 键 词：谣言检测 深度学习 多跳网络 注意力机制 多模态 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]