Fake News Recognition in social media with Multi- level Attention Fusion

Abstract

With the rapid development of social media, fake news spreads long with real information with multimodal forms, which seriously damages the credibility of media and disrupts the social order. In this case, detecting rumours effectively combined with images and texts becomes a crucial challenge that needs to be confronted. This paper proposes an end-to-end fake news recognition framework (MLAF) based on the deep neural multimodal model and multi-level attention fusion. Specifically, to fit the multimodal form of fake news, this research introduces the extra user features based on text and visual modes. To explore the interactions between different data modalities and enhance cross-modal representation, this essay first fuses the text and user features by learning an adaptive attention matrix and then further updates the text and visual modes with a multi-level attention mechanism and multimodal affine fusion. Validated by extensive experiments on the Weibo data set extracted from the real world, the results demonstrate the accuracy of the proposed model is better than existing models, and it can enhance well the cross- modal representation of fake news.