Fake News Detection Based On Multi-task Learning

With the increasing popularity of Internet,a growing number of people choose to read news from the Internet.Different from traditional news media,news from Internet has many new features,such as low cost of consumption and high timeliness.However,a large amount of news containing false information can be spread on various social media.The influx of fake news has a serious negative impact on the field of Internet news consumption.Thus,how to identify fake news and stop its dissemination in a timely manner is crucial to constructing a harmonious network atmosphere.Fake news detection is a challenging task,usually requiring annotators with domain expertise who performs careful analysis.At present,manual audit is the main way to detect fake news.However,considering the wide spread of information on the Internet and the large amount of data,the manual audit method is difficult to solve the inevitable problems such as low efficiency and high delay.With the development of artificial intelligence,many researchers expect to automatically detect fake news through artificial intelligence technology.However,the textual content of news spread on the Internet is shorter,making traditional fake news detection methods have difficulty to achieve satisfactory performance.This dissertation discusses fake news detection based on multi-task learning.News from Internet tends to be written and disseminated on various social media and covers a great quantity of topic.Therefore,the traditional machine learning method based on manual design features have difficulty to guarantee the generalization ability of design feature set.Considering news with certain topics have high probabilities to be classified as fake news,this dissertation explores the intrinsic relationship between the authenticity of news and the topic of news.This dissertation proposes a new fake news detection model based on multi-task learning(Fake news Detection via Multi-task Learning,FDML).Based on deep neural network,this model can automatically learn the corresponding features from the news content and explore the internal relationship between news authenticity and news topic.FDML can simultaneously detect the fake news and the topic of news and further improving the performance of fake news detection and topic classification.Different from traditional news,the textual content of news spread on the Internet is shorter.Moreover,fake news is generally written by an individual or group deliberately to misleading reader.It is difficult to achieve satisfactory performance in the detection of fakenews when only rely on the textual content of news.Considering that news from Internet is usually accompanied by a series of contextual information such as news speaker and credit history of speaker,such information can effectively improve the performance of fake news detection.Therefore,FDML model comprehensively considers the textual content of the news and the contextual information of the news,and further improves the performance of the fake news detection and topic classification by combining multiple features.Relevant experiments on real-word dataset demonstrate the validity of the proposed model.In addition,this dissertation implements a fake news visualization and detection system,which provides features such as data collection,data analysis,and deployment of fake news detection models.