| In recent years,social media platforms represented by Weibo and We Chat have gradually become important channels for people to obtain,share,and disseminate information due to their rich functions,convenient use,and low registration threshold.However,these conveniences have also created favorable conditions for the proliferation and dissemination of false information.The rampant spread of false information not only leads to errors in decision-making,but also seriously disrupts the public order of the platform.Therefore,the identification and governance of false information in social media has gradually attracted widespread attention from academia and industry.Currently,in the existing research on identifying fake information,the contextual semantic features and topic semantic features of text have not been fully utilized.Based on this,We propose a method for identifying fake information by integrating BERT and topic models,fully integrating the contextual semantic features and topic semantic features in the text,thereby achieving efficient identification of fake information in social media.Our main research work contents are summarized as follows:(1)Extracting Dynamic Contextual Semantic Features of Text Using BERT Pretrained Language Model.In this study,we first fine-tuned the pre-trained BERT language model on a constructed real Weibo dataset for the task of identifying fake information.Then,we input the Weibo text into the fine-tuned BERT model to obtain the dynamic contextual semantic features of the Weibo text.(2)Extracting Topic Semantic Features of Text Using BERTopic Topic Model.To extract the topic semantic features of text,Firstly,the false information is represented as document vectors using the BERT model that has been fine-tuned for the fake information identification context.Then,the dimensionality reduction algorithm UMAP is applied to reduce the dimensionality of the document vectors.Finally,the HDBSCAN clustering method is used to obtain the topic semantic features of the Weibo text.(3)Introduction of Weibo Influence and User Credibility Features from Social Context Level.In order to further improve the effectiveness of identifying false information,we introduce two feature variables,Weibo influence and user credibility,from the social context level after obtaining the dynamic contextual semantic features and topic semantic features of the text.The timeliness factor is also considered in the design of Weibo influence feature.(4)Construction and Empirical Study of a False Information Identification Model Combining BERT and Topic Modeling.The features obtained through learning are fully integrated to construct a false information identification model.Then,a real Weibo dataset is constructed as the data source for the model.A series of empirical studies,including preliminary experiments,comparative experiments,and ablation experiments,are conducted using the constructed false information identification model on the real dataset.The experimental results demonstrate that the method proposed in this paper,which combines BERT and topic modeling,has certain advantages in identifying false information on social media.The accuracy rate reaches 93.43%,which can accurately and efficiently identify false information on social media.At the same time,considering the timeliness,the influence feature of Weibo is more helpful in improving the recognition effect of false information than the influence feature of Weibo without considering timeliness. |
