Fake News Detection Based On Multi-feature Fusion

With the rapid development of the internet and online social media,people are increasingly inclined to get news from social media.However,the development of social media has also intensified the generation and spread of fake news.Especially with the COVID-19 pandemic,fake news has become more prevalent.In an era of over-reliance on social media,the spread of fake news has become a global crisis,bringing unexpected challenges to the society.Therefore,it is necessary to take measures to detect fake news and prevent its spread.Detecting fake news is so challenging that even humans can't easily tell the difference.Therefore,using deep learning technology to automatically detect fake news is of great significance to promote the benign development of journalism and maintain social stability.In view of the above problems,the main work and contributions of this thesis are as follows:Aiming at the problem that there are few datasets of Chinese long-text fake news,this thesis extends the existing dataset and proposes a fake news detection model based on multitext feature joint training.The model is composed of Max Pooling Network(MPN)and Generalized mean Pooling Network(GPN).In MPN sub-network,max-pooling is used to focus on locally most important features.In GPN sub-network,a trainable pooling layer is used to dynamically adjust the parameters,focusing on the potential multi-granularity features of text.In this model,a news article is divided into the headline and body,which are then input into MPN and GPN to extract semantic features of different levels of text.In the joint training of the two sub-networks,the maximum mean discrepancy is used to calculate the semantic consistency between the headline and the body of the news.Finally,the classification results of the two sub-networks are further fused to output the classification results of the model.The accuracy of the model on the We Chat news dataset reached 98.8%,and the F1 score reached 94.1%.For multi-modal news on the social media,the existing work tended to realize fake news detection by adding an additional subtask.However,this affects the universality and generalization ability of the model.Therefore,this thesis proposes a fake news detection model based on multi-modal feature fusion,which only takes the content of news as the main task of detection,without other additional subtasks.The model consists of four modules,which are textual feature extractor,visual feature extractor,attention mechanism and fake news detector.The textual feature extractor combines the Text-CNN and BERT to extract rich contextual information and potential key semantic information.Visual feature extractor is used to extract potential visual features of news images.The attention mechanism is used to find connections between text and image,and to weight visual features semantically similar to text.The fake news detector predicts the truth and falsehood of news based on the multi-modal feature representation learned.The model outperforms all baseline models with an accuracy of 92.2% and 83.7% on Weibo and Twitter datasets,respectively.