MRDCA: a multimodal approach for fine-grained fake news detection through integration of RoBERTa and DenseNet based upon fusion mechanism of co-attention

Being widely produced for misleading and convincing public community with biased information, various fake news has a significantly negative influence on the society as a whole. In order for effective detection of fine-grained fake news, this study developed a multimodal approach integrating RoBERTa with DenseNet through fusion mechanism of co-attention (MRDCA). RoBERTa was employed for extracting text features, and DenseNet was employed for extracting image features. The co-attention mechanism had the advantage of dynamically learning and capturing information interaction between text and image modal features. Based upon the multimodal fine-grained fake news dataset, the model of MRDCA had a higher value for all the indicators of accuracy (88.14%), macro average precision (87.16%), macro average recall (87.94%), and macro average F1 score (87.51%), comparing to unimodal approaches and other multimodal approaches through feature fusion of concatenation. More specifically, there was the unbalanced performance for MRDCA in detecting different classes of fake news. Experimental results demonstrated that the MRDCA performed better in identifying manipulated content, false connection and true than in identifying imposter content, misleading content and satire/parody. Therefore, the task of classifying samples into misleading content, imposter content and satire/parody was extremely challenging. There ought to be much room for performance promotion in detecting the three classes of fake news in future.