Research On Rumor Detection Based On Ensemble Method Of Pre-trained Language Model

With the progress and popularity of the Internet,a growing number of people utilize social media to obtain information.While social media brings convenience to people,it also brings some problems,the most notable of which is the spread of rumors.Rumors have brought many negative effects on people's lives.Although there are currently artificial rumor detection platforms,it is difficult for manual detection to respond to huge network data in a timely manner.Rumor detection becomes essential and has been actively studied in the research community.Although a lot of work has made outstanding contributions,especially the ensemble learning methods based on neural network has achieved excellent performance,but the currently proposed ensemble methods does not consider the pre-training language model as strong base learners to give full play to the performance of ensemble learning.In order to achieve the best detection results and promote the development of rumor detection,the main work of this paper is as follows:(1)We propose rumor detection model STANER(Stacking ne Twork b Ased o N BERT)that efficiently combines the pre-trained language model BERT and the Stacking algorithm.(2)We propose multi-dimensional sampling,which samples comments data as supplement for each primary learner.The strategy not only improves the performance of primary learners,but also improves diversity among primary learners.(3)We propose a de-redundacy algorithm based on BERT and DBSCAN,which can solve the redundacy problem reduced by multi-dimensional sampling.(4)We has collected a latest Chinese rumor detection dataset Weibo20,which is based on Sina Weibo platform.(5)Compared with existing SOTA methods,the STANER model has achieved better results on four real-word datasets.In addition,we also completed a series of ablation and analysis experiments to explore the STANER model and rumor detection task.