Fake News Detection Based on Attention Mechanism

Currently, most multi-modal fake news detection methods rely on fusing multi-modal features using a single criterion, overlooking the potential contribution of original multi-modal information in fake news detection and classification. This paper proposes an attention-based multi-criteria evaluation model that strengthens text feature representation on the basis of using multi-criteria. The model utilizes ResNet for extracting image feature representations and weights the output of BERT hidden layers to obtain richer text features. Subsequently, text and image features are inputted into a multi-modal feature fusion module, and the resulting output is concatenated with pooled text and image features before being passed to the final classification module, which then outputs experimental results. Experimental results demonstrate that, compared to other baseline models, the proposed model achieves better performance on Weibo and Twitter datasets.