| Rumor detection,an event-level classification problem,aims at judging the truth of rumor events.The task is to determine whether an event is a rumor based on reams of social media data.A rumor event is a collection of original posts and subsequent posts(such as comments or retweets)on social media.This dissertation studies the event-level rumor detection task from three aspects:event linear structure,propagation graph structure and topic heterogeneous graph structure.The main contents are as follows:(1)Rumor Detection with Linear Representation of Prior EventsRumor events can be expressed in various forms,among which the linear stitching of source posts and reply posts according to the publishing time is the most common.In addition,in view of the characteristics that different rumor events will discuss the same topic,this dissertation proposes a rumor detection method based on linear representation of prior events.First,the correlation between each event is calculated by cosine similarity,thus dividing the prior events for each event.Then,the two-layer attention mechanism is used to obtain the feature representation of the rumor event and the prior events.Finally,the prior features are combined to assist rumor detection.Experimental results on three real social media data sets show that this method outperforms several state-of-the-art benchmarks.(2)Rumor Detection with Event Propagation Structure and Temporal TextThe linear structure proposed in research(1)can only extract the features between text and time.In fact,there is a data structure like a multitree or graph between posts on social media.Therefore,this dissertation proposes a rumor detection method based on event propagation structure and temporal text.Firstly,the attention mechanism model is used to extract temporal text features.Secondly,each post is encoded as a feature node,and the graph model is used to extract the propagation structure features.In addition,the influence of objective information in rumor detection is enhanced.Finally,the three features are fused for rumor detection.Experimental results show that this method is superior to several state-of-the-art benchmarks.(3)Rumor Detection with Topic Heterogeneous GraphMost existing rumor detection methods are limited to events.The propagation diagram of rumors can be established through the relationship between posts within events.Meanwhile,the relationship diagram between events can be established according to the relevance of topics,and information between events can be shared by using the relationship diagram.Therefore,this dissertation proposes a rumor detection method based on topic heterogeneous graph.This method divides rumor detection into event-level and topic-level.Firstly,the event-level aims to obtain the feature representation of rumor event,which is composed of temporal feature,propagation structure feature and objective information feature.Secondly,the topic heterogeneity graph is obtained by calculating the correlation between events at the topic-level.Finally,the graph model is used to interact the information between events and complete the prediction of event categories.Experimental results show that this method is superior to several advanced benchmark models.Through the three methods,this dissertation solves some problems existing in the event level rumor detection task,so the performance of this task is improved to a certain extent. |
