Multimodal Fake News Detection Based on TinyBert Pre-Training and BiLSTM-GATV2 Models

Aiming at the current situation of large amount of news detection data and complex and variable news features, a multimodal false news detection method based on a lightweight version of pre-trained language model is proposed. Firstly, TinyBERT pretrained model is used to extract news text features; meanwhile, the propagation graph adjacency matrix is constructed to extract user propagation behavior features. Then, BiLSTM captures the forward and backward relationships of the news long text and encodes the news sequences by bi-directional information. Finally, Graph Attention Network v2 is introduced to dynamically adjust the attention weights among users and combine the textual and graph structural information for fake news prediction. The model achieves high accuracy on multiple datasets.