Research And Application Of Fake News Detection Algorithm Based On Multi-modal Feature And Semantic Enhancement

With the development of the Internet,the number of daily fake news is extremely large,so it is not enough to rely on manual to identify fake information,so the research and application of efficient fake news detection algorithm is imperative.In recent years,the form of fake news has become more and more diversified.It not only includes fake pictures and text content,but also contains external information that implies its fake features.The fake news detection model based only on the language characteristics of the fake news has many limitations:(1)The use of single-modal data such as pictures or text to detect fake news is generally not good;(2)External information with important discriminatory features is not effective used;(3)Semantic features cannot be effectively obtained for short text messages with sparse semantics;(4)The connection between the characteristics of news pictures at the physical and semantic level is not well considered.Therefore,the model that comprehensively considers the multimodal features of the text,pictures and external information of fake news has attracted the attention of industry and academia.In view of the limitations of the above fake news detection model,this paper introduces the multi-modal feature and sparse semantic enhancement module,and proposes a fake news detection algorithm model MMSEM(Multi-Modal features and Semantic Enhancement Model)based on the multi-modal feature and sparse semantic enhancement module.The feature extraction module of the MMSEM model is based on the multi-modal data of news pictures,content,headlines and external information such as news tags,authors and links,which solves the limitation of lacking integrated analysis relying on single-modal data only;MMSEM uses The TNTM(Tagging augmented neural topic model)module based on Gaussian LDA performs semantic enhancement on semantically sparse text in external information,which provides strong support for the subsequent semantic feature extraction;MMSEM uses multi-core asynchronous long convolutional neural network for news headlines and authors,and the discriminative features implied in the link information are fully extracted;MMSEM uses the attention mechanism to extract the weighted features according to the connection between the semantic and physical levels of news pictures.Finally,after fusing various modal features into the classification module,the accuracy of the model detection reached 90%after multiple iterations of training.Based on the public Weibo and Twitter datasets,the experiments were performed on single-modal data verification and multi-modal data verification.The results of the single-modal data verification experiment show that MMSEM uses the external information of the news and the semantic level information of the picture to classify better than other comparative models;the results of the multi-modal data verification experiment show that MMSEM can effectively integrate multi-modal features and the introduction of TNTM semantic enhancement module can fully extract the semantic features of the label to further improve the classification performance,and the detection accuracy is 10?20 percentage points higher than other comparison models.Based on the proposed MMSEM,combined with the current application background,a demand analysis was conducted to determine the use role and use cases of the system,then the main functions of the system were designed in detail,and finally the fake news with MMSEM as the core was realized according to the system design Prototype system tested.