Research On Weibo Rumor Detection And Stance Analysis

The spread speed and scope of news and messages on social media platforms are unprecedented,which also brings convenience to the wanton dissemination of rumors.Some targeted or deliberately fabricated rumors may provoke the emotions of the public,thereby destroying the online environment and causing terrible social impact.However,due to the massive rumors on microblog platforms,it is obviously impossible to judge the authenticity of each rumor and curb the spread of false rumors manually.Therefore,automated technologies and models are needed to analyze the rumors and further determine their truth.This paper focuses on the research of rumors on microblog platforms and mainly involves two tasks in the task system of rumor detection: stance classification and false rumor detection.The first two chapters of this paper mainly sort out the rumor detection task system and introduce the research background,related work,and widely used models or methods of these two tasks.Aiming at the shortcomings of the existing research,new solutions are proposed for both tasks respectively in the third and fourth chapters:1)For stance classification task,this paper presents a simple and effective feature template.Firstly,a complete feature template is constructed from three classic dimensions including text,user and propagation,which combines the common existing features and new features specially designed for negative categories.By testing all the features through statistical analysis and experimental analysis,18 truly effective features are extracted,and the size of the feature template is significantly reduced.On this basis,combined with traditional machine learning models,such template effectively improves the classification effect,in which Logistic Regression increases the Macro-F1 score to 57.4%.In addition,some specially designed features show good results for two categories which have a small number of samples and are difficult to detect but meaningful.2)For false rumor detection task,this paper proposes a state-independent incremental detection model.Based on the generation time of responses and retweets during the propagation of a rumor event,the model firstly introduces the Kleinberg algorithm from the field of emergency detection to identify the emergency state at each time point during the dissemination,and then segments each rumor event into consecutive sub-events according to the time points when the state changes.This ensures that the segmentation conforms to the law of rumor propagation and that the tweets within a sub-event are in the same state.Then,along the timeline of rumor propagation,the model independently trains an encoder for each sub-event,and completes the prediction after fusing its encoding with the previous sub-event's.This allows the mode to train and predict independently and incrementally on each sub-event.Experiments show that this model can significantly improve early detection accuracy and achieve further improvement with the evolution of the rumor events.