Research On Fake News Detection With Attention Mechanism And Feature Fusion

With the development of social networks,social media platforms have become a major channel for news dissemination.Due to its convenience,people are gradually seeking and consuming news through social media platforms.However,the convenience has also facilitated the rapid spread of fake news,which has a devasting impact on individuals and society.The detection and discovery of fake news have become an urgent research problem.News on social media platforms consists of multiple modalities of information,such as text content,image,and social context,which makes multimodal fake news detection a current research hotspot.Compared with the traditional unimodal fake news detection,the multimodal approach can identify fake news more effectively.Therefore,multimodal fake news detection has important research significance.Feature extraction and feature fusion are the key steps in the multimodal fake news detection task,and their effectiveness directly affects the performance of fake news detection.For the extraction of social context features,existing methods ignore correlations between social context features,and thus is difficult to obtain a more effective social context representation that is indicative of fake news.For the extraction of textual and visual features,pretrained models possess powerful processing capabilities.In feature fusion,most approaches simply concatenate a text representation representing text content and a visual representation representing images,and do not fuse multiple textual features with multiple visual features at a fine-grained level,resulting in inadequate fusion.This paper focuses on the problems in feature extraction and feature fusion for multimodal fake news detection.The main research is as follows.(1)A social context-based fake news detection model is proposed.The model uses Transformer's encoder to capture the correlations between social context features,adopts a special learnable vector to select key features that are beneficial for fake news detection from multiple social context features as a social context representation,and finally exploits the social context representation to determine the truthfulness of news.The experiment is conducted on a real-world Weibo dataset and the results show that it possesses a good capability of fake news early detection.(2)A fake news detection model based on text content and images is built.First,Ro BERTa and Swin Transformer are adopted to produce a text representation and a visual representation,respectively,which are indicative of fake news.Then,a soft attention mechanism is utilized to fuse the text representation and visual representation into a joint representation,which contains textual features and visual features.Finally,the joint representation is used to predict the category of news.The experiment is conducted on a real-world Weibo dataset and the results demonstrate the effectiveness of the model.(3)To fuse text features,visual features,and social context features,Fine-Grained Multimodal Fusion Networks(FMFN)is proposed.First,Scaled Dot-Product Attention is utilized to fuse multiple text features with multiple visual features at a fine-grained level to generate a fused text representation and a fused visual representation.Then,the fused text representation,the fused visual representation,and the social context representation are transformed into a joint representation containing the information of the three modalities using a soft attention mechanism.Finally,this joint representation is used for the identification and discovery of fake news.FMFN achieves an accuracy of 91.0% on a real-world Weibo dataset,which strongly demonstrates the effectiveness of FMFN.