Research On Rumor Detection Technology Based On Deep Learning Model

The rapid development of social media has changed the main ways for the public to obtain or disseminate information.Online social platforms such as Sina Weibo have become the main tools for people's information exchange due to their ease of use,freedom of expression,and rapid dissemination.But these characteristics also lead to the proliferation of rumors on social platforms.The spread of rumors seriously violates the vital interests of citizens and disrupts the public order of the Internet.The method of manually detecting rumors is costly and inefficient,so automatic rumors detection technology has become an important research direction.At present,automatic rumor detection technology mainly has two types of methods,one is the rumor detection method based on manual features,and the other is the rumor detection method based on deep learning.The methods based on manual features rely too much on prior knowledge when extracting features.And their expressive ability is limited.So the detection result is not very satisfactory.The rumor detection methods based on deep learning can automatically learn potential features beneficial to the task from the data.So the methods significantly improve the detection performance of the model.For this reason,the application of deep learning to automatic rumor detection tasks has become the mainstream of current technology.Rumor detection task can be divided into message-level rumor detection task and event-level rumor detection task.In the message-level rumor detection task,the message content includes not only text but also images.Text and image complement each other when describing the same thing.They describe the same thing from different perspectives respectively.However,most of the previous detection methods based on deep learning only focus on the content in the text or image in isolation.So,some researchers proposed some detection methods combining the two modes information.However,the existing combination methods do not make full use of the correlation between image and text.And they ignore the difference between the low content consistency between image and text in rumors and the high content consistency between image and text in non-rumors.In the event-level rumor detection task,faced with a suspicious micro blog,people tend to make comments to provide various clues to question its authenticity.Therefore,compared with the real events,the retweets and comments of rumor events contain more doubts,negation and other stance.The semantic expression of each message is often affected by the information released before it.Previous deep learning-based detection methods generally used recursive neural networks to learn the time series relationship in the event and extract the stance information in the text.But these methods still have some shortcomings when extracting stance information.The specific questions are shown as follows.When mining the attitudes and opinions in the message,the influence of user credibility is ignored.And the problem that the network cannot learn the context dependence effectively due to the long sequence of events is not solved.In addition to differences in attitude,viewpoints and other stance information,since real events are real,users tend to publish evidential information learned from other sources in the retweets and comments.The information is as a supplement to the content of the source micro blog to prove the authenticity of the event.So,compared with rumor events,there is more supplementary information to the content of the source micro blog in the retweets and comments.But the existing event-level methods do not combine this distinction to detect rumors.To this end,This thesis focuses on how to use deep learning models to extract this potential information to further improve the accuracy of rumor detection.The specific research content is as follows:1.Aiming at the task of message-level rumor detection,this thesis distinguishes rumors from non-rumors by mining the content consistency between image and text.In order to calculate the consistency,a simple method is to directly extract the global feature representation of image and text and project them into a common feature space for calculation.However,the consistency obtained by this coarse-grained method often cannot accurately distinguish rumors from non-rumors.So this thesis makes a fine-grained comparison.To achieve this goal,this thesis constructs a bidirectional attention alignment model.First,the image is divided into visual region elements,and the text is divided into word elements.Then,the bidirectional attention mechanism is used to match each element to the closest semantic element in another model to form a pair of elements.Then the elements of different models are projected into the common semantic feature space and the consistency between the elements is calculated.Finally,the consistency of all element pairs is summarized for classification.Experiments show that the model can effectively distinguish rumors early in the release.And they also show the model plays an important role in the early warning of rumors.2.Aiming at the task of event-level rumor detection,this thesis combines stance information and content supplementary information to improve the detection effect.In order to these two kinds of information,this thesis constructs a multi-information fusion model.The model is divided into two modules,one is stance information mining module,the other is content supplementary calculation module.In the stance information mining module,because there are too many messages in some hot events,it is impossible for the model to learn the interdependence between messages.Therefore,this paper divides the events into a series of sub-events according to time,forming the message layer and sub-event layer.First,we learn the dependencies between messages within each sub-event at the message level.Secondly,since messages in different time stages contribute differently to detection results,and messages published by users with different credibility contribute differently to detection results,the self-attention mechanism combined with user characteristics is used to assign different weights to messages.Then the representations of all messages are aggregated to obtain the text representations of sub-events.Thirdly,the dependency between sub-events at the sub-event level is learned.And the self-attention mechanism is used to assign different weights to each sub-event.Finally,the text representation of the event is obtained by aggregating the representation of the sub-events.The representation contains stance information for detection.In the content supplementary calculation module,firstly,the respective keywords of the source micro blog,the retweets and the comments are extracted.Then the number of newly appearing keywords in the retweets and comments is used as the content supplementary.This is because the keyword contains the main content information.The more new keywords shows that information degree is higher.Finally,all the supplementary degrees are combined into a vector as the overall supplementary degree representation.The representations obtained by the two modules are combined for classification.Experiments show the model can detect rumors more effectively than other benchmark models.