Fake News Detection Based On Graph Convolutional Networks And Prompt Tuning Learning

With the vigorous development of social networks,social media platforms have not only become the main channel for news dissemination,but also promoted the rapid spread of false news,which has brought serious harm to individuals and society.Therefore,fake news needs to be identified and dealt with quickly and efficiently.At present,the main method of news detection relies on the manual operation of the auditors of the platform,and the labor cost remains high.Researchers have tried to use news detection algorithms for intelligent identification,but due to commercial barriers and personal privacy protection,the news data required by many advanced algorithms cannot be easily obtained.This paper focuses on the detection of fake news in the case of limited types of news data labels and insufficient samples.The main research content of this paper will be carried out from three aspects:(1)In the case of difficult access to complex news information and limited types of labeled data,this paper proposes a new model GCA based on graph convolutional neural network and pre-trained language model ALBERT.The model encodes the global semantic information contained in the text sequence through the graph convolutional network based on word cooccurrence information built on the vocabulary graph,and then feeds the dual embeddings containing graphs and words to the self-attention of the ALBERT encoder in the force layer.Word embeddings and graph embeddings interact through a self-attention mechanism.The final classifier not only obtains the local information and global information of the text content,but also uses the attention mechanism to combine the two.The final classification representation can integrate the local information and global information of the news text,and fully tap the information potential of the news text.,enabling the model to achieve excellent fake news detection results only relying on news text information.(2)Aiming at the problems of difficult access to news data and insufficient training samples,this paper designs a suitable prompt template through a pre-trained large-scale textto-text converter based on the method of prompt learning and combines it with GCA to generate a GCAprompt model.Based on the function of the prompt template,GCAprompt can transform the prediction and classification task of news text into a specific task of predicting words,thereby bridging the gap between downstream tasks and pre-training methods,enabling the model to directly extract relevant knowledge from the pre-training process,Only a small amount of news text data is needed for fine-tuning training to efficiently identify fake news.(3)The GCAprompt model is tested on two real news datasets.The results show that the performance of the GCAprompt model is better than other baseline methods under the condition of using only news text data.At the same time,GCAprompt is still effective in a small sample data environment,and only needs 100 news texts to achieve efficient false news detection.In summary,through a series of experiments,it is proved that the GCAprompt model proposed in this paper can effectively solve the problem of false news detection and provide technical support for the public opinion supervision department to purify the network space environment.