基于多模态特征融合的社交媒体虚假新闻检测

Fake News Detection Based on Multimodal Feature Fusion on Social Media

【目的/意义】社交媒体在改变新闻传播以及人类获取信息方式的同时,也成为了虚假新闻传播的主要渠道。因此,快速识别社交媒体中的虚假新闻,扼制虚假信息的传播,对净化网络空间、维护公共安全至关重要。【方法/过程】为了有效识别社交媒体上发布的虚假新闻,本文基于对虚假新闻内容特征的深入剖析,分别设计了文本词向量、文本情感、图像底层、图像语义特征的表示方法,用以提取社交网络中虚假新闻的图像特征信息和文本特征信息,构建多模态特征融合的虚假新闻检测模型,并使用MediaEval2015数据集对模型性能进行效果验证。【结果/结论】通过对比分析不同特征组合方式和不同分类方法的实验结果,发现融合文本特征和图像特征的多模态模型可以有效提升虚假新闻检测效果。【创新/局限】研究从多模态的角度设计了虚假新闻检测模型,融合了文本与图像的多种特征。然而采用向量拼接来实现特征融合,不仅无法实现各种特征的充分互补,而且容易造成维度灾难。

【Purpose/significance】Social media has not only changed the way news spreads and the way people get news,but also become the main channel of fake news dissemination,which causes great harm to society.【Method/process】In order to effectively identify the fake news published on social media,this article designs the text word embedding,text sentiment,image bottom layer,and image semantic feature representation methods to extract the image features in the tweet based on the analysis of the characteristics of the fake news content information and text feature information,a fake news detection model of multimodal feature fusion was constructed,and the performance of the model was verified using the MediaEval2015 dataset.【Result/conclusion】By comparing and analyzing the experimental results of different feature combinations,it is found that the multimodal model combining text features and image features can effectively improve the effect of fake news detection.【Innovation/limitation】A fake news detection model has been designed from the perspective of multimodality,which combines various features of text and image.However,using vector mosaic to realize feature fusion can’t achieve full complementarity of various features and easily cause dimension disaster.