Zeng Jiangfeng, Zhang Yin, Ma Xiao
School of Information Management, Central China Normal University, Wuhan, China.
School of Information and Communication Engineering, University of Electronic Science and Technology of China, Shenzhen, China.
Sustain Cities Soc. 2021 Mar;66:102652. doi: 10.1016/j.scs.2020.102652. Epub 2020 Dec 14.
In recent years, major emergencies have occurred frequently all over the world. When a major global public heath emergency like COVID-19 broke out, an increasing number of fake news in social media networks are exposed to the public. Automatically detecting the veracity of a news article ensures people receive truthful information, which is beneficial to the epidemic prevention and control. However, most of the existing fake news detection methods focus on inferring clues from text-only content, which ignores the semantic correlations across multimodalities. In this work, we propose a novel approach for Fake News Detection by comprehensively mining the Semantic Correlations between Text content and Images attached (FND-SCTI). First, we learn image representations via the pretrained VGG model, and use them to enhance the learning of text representation via hierarchical attention mechanism. Second, a multimodal variational autoencoder is exploited to learn a fused representation of textual and visual content. Third, the image-enhanced text representation and the multimodal fusion eigenvector are combined to train the fake news detector. Experimental results on two real-world fake news datasets, Twitter and Weibo, demonstrate that our model outperforms seven competitive approaches, and is able to capture the semantic correlations among multimodal contents.
近年来，重大突发事件在全球频繁发生。当像新冠疫情这样的重大全球公共卫生突发事件爆发时，社交媒体网络中越来越多的虚假新闻被公众知晓。自动检测新闻文章的真实性可确保人们获得真实信息，这有利于疫情防控。然而，现有的大多数虚假新闻检测方法都集中于仅从文本内容中推断线索，这忽略了多模态之间的语义关联。在这项工作中，我们提出了一种通过全面挖掘文本内容与所附图像之间的语义关联来进行虚假新闻检测的新方法（FND-SCTI）。首先，我们通过预训练的VGG模型学习图像表示，并利用它们通过分层注意力机制增强文本表示的学习。其次，利用多模态变分自编码器学习文本和视觉内容的融合表示。第三，将图像增强的文本表示和多模态融合特征向量相结合来训练虚假新闻检测器。在两个真实世界的虚假新闻数据集Twitter和微博上的实验结果表明，我们的模型优于七种有竞争力的方法，并且能够捕捉多模态内容之间的语义关联。
