Research And Implementation Of Multi Model Fake News Classification System Based On Bert

With the development of mobile Internet,the emergence of We Chat,Weibo and other network social platforms has changed the way people get news.The fast news transmission speed on network platforms and the weak supervision compared with traditional platforms have led to the greater harm of fake news on network platforms.Due to the rapid spread of fake news,the method of manual screening can not stop its spread in time.Therefore,in recent years,automatic detection of fake news has become a research hotspot.Many existing research methods on fake news need text features,information transmission features and other features to classify.However,in the early detection of fake news,there is often only text feature,and information transmission feature and user feedback feature cannot be obtained effectively.In order to solve the problem of early detection of fake news,this paper proposes a multimodel fusion detection method based on BERT to analyze the text content of news and detect fake news.The algorithm research in this paper is mainly divided into two parts.Firstly,through studying the basic knowledge of text vectorized representation,a kind of model combining word embedding based on BERT model and word embedding based on Word2 Vec model has been proposed,expand the meaning of the text vector from multiple dimensions,make the subsequent model learn more deep semantic representation,get better detection efficiency.Secondly,this paper has carried on the research of classification model fusion.For analyze the text features with different emphasis,Text CNN,Bi LSTM and Transformer model has been selceted to be the basic model of the fusion model.And based on these models proposed a model combined Text CNN model and Bi LSTM model and the Self-Attention mechanism,named SA-BLTCNN model.This model combines the characteristics that CNN is good at extracting local features and Bi LSTM is good at extracting long-distance features,and uses Self-Attention mechanism to strengthen the feature extraction of keywords in texts,which can dig out more deep semantic information of news,thus improving the detection accuracy of fake news.Through the comparative experiment in this paper,it is proved that it can effectively improve the detection accuracy of fake news.Then the model fusion was carried out on the four basic models,and different fake news detection fusion models were obtained,including simple voting fusion model,Stacking fusion model and Adaboost fusion model.Through the method of model fusion,combined with the characteristics of each basic model,the model finally obtained has a higher accuracy and stability.After that,three different fusion models were experimentally analyzed and their effectiveness was further verified.At the end of the paper,a system suitable for early automatic detection of fake news is designed based on the multi-model fusion fake news detection model obtained in the previous stage,and the modules of fake news detection and fake news retrieval have been implemented.Finally,the relevant results have been shown.