摘要
为解决多模态虚假新闻检测在特征融合过程中无法捕捉不同模态之间复杂关系的问题,提出了基于多级融合网络的多模态虚假新闻检测模型。该模型通过特征提取器提取不同模态的特征,采用多级融合网络进行特征融合,该网络通过跨模态交互层来探究不同模态间的信息交互,通过拼接融合层避免融合过程中的维度灾难,经过两层融合层后生成最终融合特征;最后将融合后的多模态特征通过新闻检测器判断新闻真假。该模型在Weibo数据集上取得92.0%的准确率,较最优的基线模型提升了将近5.1%,验证了该模型能够有效的捕获不同模态间的依赖关系,深度融合不同模态特征,从而提高虚假新闻检测的准确率。
To address the limitation in capturing complex inter-modal relationships during feature fusion for multimodal fake news detection,a multimodal fake news detection model based on a multi-level fusion network was proposed.Features from different modalities were extracted using feature extractors and fused through a multi-level fusion network.Cross-modal interactions were captured by the cross-modal interaction layers,while the curse of dimensionality during fusion was avoided via the concatenation fusion layer.After two fusion layers,the final fused features were generated and subsequently fed into the fake news detector for authenticity prediction.Experimental results demonstrate that the proposed model achieves an accuracy of 92.0%on the Weibo dataset,outperforming the state-of-the-art baseline by nearly 5.1%,validating that this model can effectively capture the dependencies between different modalities and deeply fuse multimodal features,thereby improving the accuracy of fake news detection.
作者
夏旭
王晓强
齐承睿
陈新一
李少波
庄旭菲
XIA Xu;WANG Xiao-qiang;QI Cheng-rui;CHEN Xin-yi;LI Shao-bo;ZHUANG Xu-fei(School of Information Engineering,Inner Mongolia University of Technology,Hohhot 010080,China)
出处
《计算机工程与设计》
北大核心
2025年第10期2866-2872,共7页
Computer Engineering and Design
基金
内蒙古自然科学基金项目(2023MS06021)
国家自然科学基金项目(62165011)
2023年自治区直属高校基本科研业务费基金项目(JY20230065)。
关键词
虚假新闻检测
多模态分析
多模态特征融合
多级融合
情感特征
社交媒体
深度学习
fake news detection
multimodal analysis
multimodal feature fusion
multi-level fusion
affective feature
social media
deep learning
