Research On Rumor Detection Method Based On Generative Adversarial Network

With the rapid development of the Internet,various news websites,multimedia clients,and public accounts hosted by party and government departments have become emerging information media.Compared with traditional media such as TV,newspapers and periodicals,people prefer to get news on these emerging media.These news media platforms have brought people into the era of big data information with their explosive information iteration speed and amazing dissemination speed.Due to the rapid update and dissemination of information and the people's high reliance on emerging media,which requires greater supervision of online information and blocking the spread of online rumors.Rumors not only affect the behavior that misleads people,but also harm the development of society.Now,there are network police on the Internet to manually review information,but this method of rumor detection is extremely inefficient.Therefore,we need to design an end-to-end algorithm to effectively detect rumors and improve the ability to monitor network information.At present,scholars have proposed some rumor detection methods,such as SVM based on traditional machine learning algorithms,neural network GRU,which have achieved good results in rumor datasets.But these algorithms essentially ignore the state of information dissemination.Internet rumors will artificially add,delete,and replace information during the spreading process,which is confusing.An Internet rumor often has different versions at different time periods,which poses a challenge to the existing rumor detection algorithm.The extraction of key features of rumors and the improvement of the generalization ability of rumors detection models are the key points that need to be solved at present.To resolve the above issues,this dissertation launched the following main research.The main work of this dissertation is as follows:(1)A text rumor detection method based on a generative confrontation network have put forward in this paper.The model consists of a text classifier and a text generator.The text generator uses the SeqToSeq model to realize the sequence generation of the rumor text,simulating the feature modification or addition and deletion of the rumor in the propagation process in real life,and better improving the identification ability of the rumor text classifier.The text classifier adopts a long and short-term memory network structure with attention,which can effectively extract and classify rumor texts of different lengths.(2)A self-attention mechanism-based generation of anti-web text rumors detection method is proposed.The SeqToSeq model is currently mainly used for natural language translation tasks.In the rumor detection model,it is necessary to perform feature extraction and text generation of the rumor text.Compared with the SeqToSeq model,the Transformer network model has obvious advantages in feature extraction.The multi-head self-attention mechanism can better extract the relationship of words to words and words to sentences.Therefore,the text rumor detection model transforms the text generator into a network model based on the self-attention structure,which can better extract text features and generate approximate samples to strengthen the rumor classification model and improve the generalization ability of the text rumor detection model.(3)Further,verifies the effect and robustness of the self-attention mechanism-based generative adversarial network text rumor detection method on early rumor detection.Comparing the detection accuracy of this method with other methods,the experimental results verify that this method performs better than other methods in the early rumor detection task.(4)Further,verifies the effect of self-attention mechanism on the generated sequence.This paper uses Sina Weibo and Twitter as the research platform to compare the sequence generated by the self-attention mechanism with the SeqToSeq network to further confirm the reliability of the model.Meanwhile,the self-attention structure also improves the rumor detection ability of the text rumor model...