Research On Social Network Rumor Detection And Harmfulness Prediction

Online social network has become an important information releasing and sharing platform.These platforms have relatively low thresholds for publishing information,with diversified information and liberalized expression,making them an excellent platform for the generation and spread of rumors.Online rumors are often harmful to a certain extent and easily lead to social panic,which is particularly harmful to social progress and national development.The purpose of rumor detection technology is to quickly locate network information suspected of rumors,and it is the basic work of network rumor management.In this dissertation focuses on solving the key roblems faced by automatic rumor detection.Firstly,the lack of key information background: social network information is short in content and wide in scope,and usually lacks background knowledge of keywords,which can easily lead to misunderstandings and rumors.Secondly,the difference in the transfer sample domain:the multi-modal early rumor detection has few training samples,but the source domain data used by the pre-training model and the social network rumor data are different,and the direct transfer effect is not good.Thirdly,the detection results are of limited use: most of the existing methods remain in the binary classification stage,without discussing which more fine-grained category they belong to,and cannot give the measurement of their harmfulness,making it difficult to narrow the scope for investigation and evidence collection.This dissertation focuses on the following three aspects:(1)For the lack of key information background in social network short texts,a rumor detection method based on entity recognition and sentence reconstruction is proposed,focusing on improving the performance of rumor detection from the perspective of text representation optimization.The method first uses entity recognition and online knowledge base to explain and embed the text to be tested,and initially understand the key entities in it.Then,a sentence reconstruction method is designed to adjust the word order and word frequency of sentences,and reduce the difficulty of machine understanding of text content through semantic enhancement.Next,build a feature map containing comment forwarding data and statistical features.The statistical features mainly include three aspects: the language features of the rumor content,the user features involved in the rumor,and the structural features of the dissemination network.Finally,a deep learning classifier is used to classify the feature maps to implement the rumor detection function.Experiments on Twitter and Weibo datasets demonstrate that the proposed method achieves better performance than previous related works.(2)For the domain gap between the social network rumor data and the pre-training model source domain data,a multimodal early rumor detection method based on domain adaptation is proposed to reduce the domain gap and improve the detection performance.The method consists of three parts: a textual feature extractor,a visual feature extractor,and a fusion and classification network.In the text feature extractor part,to improve the diversity and stability of text representation,multi-task sharing layer,task-specific encoder and selection layer are applied to build a network-based domain adaptation model.In the visual feature extractor part,the domain gap between social network pictures and image pre-training datasets is narrowed by adversarial domain adaptation methods.Finally,the fusion and classification network uses two feature fusion strategies,feature-level fusion and decision-level fusion,to synthesize image and text features.Experiments on multimodal datasets show that the proposed domain adaptation-based multimodal network outperforms related works.(3)In view of the limited effect of rumor detection results on subsequent rumor refutation work,a multi-class method and rumor harmfulness measurement method are proposed to detect rumors in more detail.First,by labeling the existing rumor data set,four types of labels are obtained: content,source,cause and writing.Then,a classifier based on stable learning is designed to improve the performance on unbalanced datasets.The experiments of three classification strategies on multi-class,multi-label and multi-task show that the combination of the stable learning module and the multi-task mode achieves the best result.On the other hand,this study proposes two harmfulness measurements based on public opinion index and manual annotation,and trains corresponding regression models.The interaction between rumor multiclass and harmfulness prediction task relies on cross-fusion method.The research on multiclass and harmfulness prediction of rumors expands the traditional rumor detection work.