Research On Disinformation Detection Based On Fusion Of Event And Stance Analysis

In today’s highly advanced internet era,social media has become the primary source for people to obtain information and express their emotions.The development of social media has brought great convenience to people’s daily lives and communication,but it has also provided a medium for the proliferation and dissemination of disinformation.Common types of disinformation found on the internet include fake news,rumours,misleading advertisements,and click fraud.Among these,fake news and rumours are the two most widely spread and harmful types of disinformation.To mitigate and prevent the harms caused by fake information,it is crucial to have real-time monitoring and analysis of fake information on the internet.However,the current state of fake information detection still has some limitations.Firstly,the role of event information has not been adequately emphasized.Current detection models mainly extract features from the target text and its immediate context,but they lack a deep utilization of event information.Secondly,the utilization of stance information is still insufficient.Stance information often reflects collective cognition and plays a significant role in detecting fake information.Lastly,the integration of information from external news and social media comments is not fully explored.To identify fake information,it is essential to leverage the background and evidence provided by external news sources and also tap into the collective knowledge and opinions present in social media comments.To address the above-mentioned issues,this paper investigates the following areas:Firstly,in response to the insufficient use of event information for disinformation detection: This paper proposes a method for identifying fake news confirmation relationships that integrate news event information based on event analysis.This method can analyze events in news,obtain the background and supporting or opposing evidence for false information,and judge the credibility of the information to be detected.Specifically,the model obtains the context and event information in both the data to be detected and relevant news through event and text encoding.The information fusion module filters out irrelevant events and highlights important events based on the attention mechanism for the possible appearance of multiple events in the news.Experimental results on public datasets demonstrate that the proposed model can effectively utilize event information in the text to improve disinformation detection performance.Secondly,in response to the insufficient use of stance information for disinformation detection: This paper designs a joint model for rumor and stance detection based on the Partition Filter Network(PFN),which can leverage shared and interactive features for both tasks.In addition,this paper constructs a stance network in dialogues based on the associational relationships of social media comments and applies the Graph Transformer to utilize collective interaction features in the stance network.Experiments show that the model can effectively utilize stance information in social media comments to improve the disinformation detection effect in social networks.Thirdly,in response to the inadequate integration of external news and social media comments: This paper first constructs an event graph for a rumor dataset through event extraction and event relation extraction,which presents objective information in external news.Later,this paper obtains group knowledge and opinions of social network users through the stance network.Based on the constructed event graph and stance network,this paper achieves effective integration of external objective information and subjective information,improving the identification of disinformation in social networks.Fourthly,to overcome the lack of global semantic information in current document-level event extraction methods: This paper proposes an extension of the sentence-level Abstract Meaning Representation(AMR)structure to the documentlevel,incorporating document-level information like coreference chains and maximallength noun phrase chains.Based on this enriched semantic information,the paper proposes a document-level event extraction model that adopts document-level Abstract Meaning Representation(AMR).The experimental results on three different documentlevel event extraction datasets demonstrate that the proposed model achieves good performance in document-level event extraction tasks,outperforming existing models in distant argument recognition and handling long document data.Fifthly,to address the problem that current event coreference research ignores complex event coreference: This paper provides a discourse deixis annotation dataset for complex events and designs a multi-granular graph network model based on this dataset,which integrates different granularity semantic information such as characters,words,phrases,and sentences to identify event coreference in discourse.Experimental results show that the proposed annotation corpus can support the training of large-scale deep learning models,and the proposed model achieves good performance on this corpus.This paper is positioned to utilize natural language processing techniques for detecting fake information.With this goal in mind,the paper conducts research on event analysis,stance analysis,and fake information detection based on event and stance information.It introduces new models and methods for these tasks,expands on existing research,and validates the effectiveness of the proposed methods through experiments.