Research On Rumor Detection Method Via On Knowledge Distillation

With the rapid popularization of 5G and the rapid development of new media technology,the speed of information dissemination has become increasingly rapid.Online social media has become the main way for most people to obtain news and share opinions.However,while social media facilitates our access to information,it also provides a breeding ground for the widespread dissemination of rumors.Through social media,rumors can spread widely in a short period of time,posing a huge threat to people’s property security and social stability.Therefore,how to automatically and efficiently detect rumors on social media and mitigate the harm caused by rumors has become an urgent problem in both the academic and industrial communities.In this paper,we take Weibo as the research scenario,and uses Chinese blogs as the basic data to carry out related work on rumor detection.The main research purpose of this paper is to mine rumor-related features from Weibo data,then design a rumor detection model based on deep learning methods and knowledge distillation technology,and verify the feasibility of the model through experiments.The main work content of this paper is as follows:(1)A new rumor detection dataset weibo22 was collected and constructed,and multi-dimensional data analysis experiment was carried out on it.Considering that there are differences between the existing rumor detection data set in the Chinese scene and the data on the current social platform,such as the length of the original post text,the distribution of event topics,and the source of the post,these differences may lead to the model trained on the public data has good inference ability but the performance is not good when it is used for current social platform detection.In order to solve the above problems,this paper takes Weibo as the research scenario,constructs a new rumor detection dataset in the Chinese scene based on Sina Weibo,names it the weibo22 dataset,and conducts an analysis of the text features and statistical features of the weibo22 dataset.(2)A rumor detection algorithm based on hierarchical gated interactive fusion network is designed.Based on the pre-trained language model BERT and related technology of feature fusion,this paper designs a rumor detection algorithm based on hierarchical gated interactive fusion network.The algorithm uses the pre-trained language model BERT to semantically encode the original posts and comments of Weibo messages to obtain the semantic features of the original posts and comments;at the same time,it extracts the emotional features of the original posts and comments with the help of sentiment analysis.Subsequently,the designed gating unit is used to mine the correlation between semantic features and emotional features to achieve deep fusion of features,and the fused features are input into the classifier for prediction and classification.Experiments were carried out on the public weibo dataset and the self-built weibo22 dataset in this paper.The detection accuracy rates reached 96.71%and 97.36%,respectively,compared with the best baseline model,which increased by 0.84%and 1.31%,respectively.(3)A lightweight rumor detection model algorithm is studied.Considering that the high-performance rumor detection model built has a large amount of parameters,if it is deployed to the system,it requires high hardware resources,and the inference speed after deployment cannot meet the actual needs.In order to solve the above problems,based on the relevant technology of knowledge distillation,this paper performs lightweight operations on the high-performance model proposed above to obtain a lightweight rumor detection model.Compared with the original model,the parameters of this model are reduced by 97%,the speed is increased by 72 times,and has the 95%prediction accuracy of the original model.(4)A prototype system for rumor detection is designed and implemented.By analyzing the deficiencies of existing rumor refuting platforms,this paper designs and implements a prototype system for rumor detection.The system has the functions of data acquisition,data viewing,feature viewing and model prediction.Among them,the data collection module can be used to collect new data and store it in the database;after that,the existing data in the database can be viewed through the data viewing module;the statistical characteristics of rumor data and non-rumor data can be understood through the feature viewing module;the model prediction function provides detection Algorithms are used for rumor detection.