Research On Rumor Detection Algorithms In Multimodal Social Media

The popularity of social media has completely changed the way that people obtain information.More and more people choose to express their feelings and share their lives by social media.Unfortunately,due to a large number of users do not carefully verify the released contents when they post information and share their opinions,various rumors have been fostered on social media websites.The widespread spread of these rumors will bring new threats to political,economic cultural fields and affect people's normal life.In order to strengthen the rumor detection and prevent the spread of rumors,many approaches have been proposed to detect rumors.The early rumor detection platform(e.g.,snopes.com)mainly reported through users,and then invited experts or institutions in related fields to confirm.Although these methods can achieve the purpose of rumor detection,the timeliness of detection has obvious limitations.Thus,how to detect rumors automatically has become a key research direction in recent years.To date,many automatic detection approaches have been proposed to improve the efficiency of rumor detection,but these methods may suffer from the following limitations:(1)At the semantic level,most previous models are based on convolutional neural network or recurrent neural network to obtain the semantic features of posts.However,these models cannot obtain the long-distance dependence relationship between words when facing long text content,which makes the semantic features incomplete;(2)In post level,many existing methods only consider the text content,usually ignoring the multi-modal information in the posts;(3)In event level,existing approaches typically only use the temporal sequence model to capture temporal feature of events,the local and global information has not been well investigated yet.To overcome these limitations,this thesis proposes two social media rumor detection methods:(1)A rumor detection method is based on the BERT-LSTM network,which uses a pre-trained BERT model to extract textual representations that contain rich semantic features from the text content of posts,BERT-LSTM network solves the problem of long-distance dependence on semantic features.Subsequently,the text representation of the posts will be sent to the LSTM network to further extract event features for rumor detection.Experimental results on two rumor detection data sets show that the BERTLSTM network can effectively capture long-distance dependencies between words,improving the accuracy of rumor detection;(2)A rumor detection method is based on the multi-modal multi-level event network(MMEN).In terms of post features,MMEN has designed a multi-modal feature extraction module to extract text content and visual content;in terms of event features,the multi-level event network proposed by MMEN utilizes mean pooling,recurrent neural networks,and convolutional neural networks to capture global,temporal and local information of each event,and send the comprehensive representation of the event to the rumor detection layer for rumor event detection.Experimental results on two rumor detection data sets show that MMEN has an accuracy rate,which is 4% higher than the existing benchmark model.