融合语义聚类与Haar小波频域的多模态虚假新闻检测方法OA
Multimodal False News Detection Method Based on Semantic Clustering and Haar Wavelet Frequency Domain
现有多模态虚假新闻检测方法对文本信息多从全局语义角度分析,忽视了局部语义之间的不一致性,导致模态间融合不充分、相关性较低,对图像多层次的频域特征以及嵌入文本信息的捕捉利用存在局限,导致模型难以挖掘图像的潜在语义特征.针对上述问题,提出了一种融合语义聚类与Haar小波频域特征的多模态虚假新闻检测方法(SC-HWFF-MFD).该方法构建面向多情境差异化的语义表征空间,并设计无监督情境语义对齐优化阶段,以增强新闻局部语义特征与对应情境之间的语义一致性.利用Haar小波变换实现图像高频的细节特征(如纹理和边缘)以及低频的背景和结构等信息的层次化建模,引入多尺度卷积和注意力机制,实现视觉特征空间域-频域的深层次关联与交互,提升视觉特征的表示能力.为解决图文特征之间的失配问题,设计了图文匹配融合策略,该策略以图像文本作为辅助信息,并结合注意力融合模块逐步实现图像与文本的对齐融合.实验结果表明,SC-HWFF-MFD方法在Weibo数据集和Twitter数据集上分类准确率Acc、虚假新闻F1值和真实新闻F1值分别为92.2%、91.8%、91.5%和84.9%、83.7%、84.9%,均优于现有的基线模型,证实了该方法在虚假新闻鉴别中的有效性.
The existing multimodal fake news detection methods mostly analyze the text information from the global semantic perspective,ignoring the inconsistency between local semantics,resulting in insufficient fusion and low correlation between modalities.There are limitations in the capture and utilization of the multi-level frequency domain features of the image and the embedded text information,which makes the model difficult to mine the potential semantic features of the image.To solve the above problems,this paper proposes a multimodal fake news detection method based on semantic clustering and Haar wavelet frequency domain features(SC-HWFF-MFD).Firstly,a semantic representation space for multi-context differentiation is constructed,and an unsupervised context semantic alignment optimization stage is designed to enhance the semantic consistency between the local semantic features of news and the corresponding context.Secondly,the Haar wavelet transform is used to realize the hierarchical modeling of the high-frequency detail features(such as texture and edge)and the low-frequency background and structure information of the image.The multi-scale convolution and attention mechanism are introduced to realize the deep correlation and interaction between the spatial domain and frequency domain of visual features,and improve the representation ability of visual features.In addition,in order to solve the problem of mismatch between image and text features,an image-text matching fusion strategy is designed.The strategy uses the image text as auxiliary information and combines the attention fusion module to gradually realize the alignment and fusion of the image and text.Experimental results show that the classification accuracy Acc,F1 of fake news and F1 of real news of SC-HWFF-MFD method on Weibo dataset and Twitter dataset are 92.2%,91.8%,91.5%,84.9%,83.7%and 84.9%,respectively,which are better than the existing baseline models.The effectiveness of the proposed method in fake news identification is confirmed.
杨力;廖远
西南石油大学 计算机与软件学院,成都 610500西南石油大学 计算机与软件学院,成都 610500
信息技术与安全科学
多模态虚假新闻检测语义聚类Haar小波频域图文匹配融合注意力机制
multimodal false news detectionsemantic clusteringHaar wavelet frequency domainimage-text matching fusionattention mechanism
《计算机科学与探索》 2026 (5)
1431-1442,12
国家自然科学基金(61175122)四川省科技计划项目(2022NSFSC0555).This work was supported by the National Natural Science Foundation of China(61175122),and the Science and Technology Program of Sichuan Province(2022NSFSC0555).
评论