Microblog Rumor Detection Research Based On BERT

As an excellent Internet product serving social networking,Microblog has shaped the pattern of decentralized information transmission and is the main carrier of network public opinion generation.As a kind of Internet rumors,microblog rumors break through the limitations of traditional media in space,have a faster spread speed and a wider spread range,and cause greater influence and harm.Therefore,how to timely and accurately detect microblog rumors has become a hot research topic in recent years.A large number of rumour detection studies have been conducted on Twitter and Sina Weibo at home and abroad.Although the rumour detection method based on deep learning can automatically learn the potential features in the text,it still uses the word vector technology represented by Word2 Vec in the text representation.These technologies have the problem of polysemy,and the performance improvement for subsequent tasks is very limited.At the same time,a considerable number of researches on microblog rumor detection aims at extracting effective features from microblog events,which cannot meet the needs of early microblog rumor detection.In order to solve the existing problems in the research of microblog rumor detection,this paper proposes a microblog rumor detection method based on the pre-trained language model BERT,BERT can effectively deal with the polysemy problem and has a significant improvement in the effect of several typical downstream tasks in natural language processing.The following two tasks are mainly carried out in this paper:The first work is to realize the early detection of microblog rumors based on BERT.Inspired by the fine-tuning model method and the use of pre-training models to complete specific tasks,this paper proposes an early detection model for microblog rumors based on BERT,BERT-CNN and BERT-GRU.The input of the model is the text of the source microblog,and the text representation and feature extraction of the source microblog can be carried out as soon as the microblog is published,which can meet the needs of early detection.Experimental results show that the BERT-GRU microblog rumor early detection model has an accuracy rate of 94.6%,which is 11.8% higher than the Word2vec-GRU model,which proves that BERT is better than Word2 vec in improving the performance of microblog rumor detection task and text representation.The second work is to implement BERT-based microblog rumor event detection.According to previous research,microblog events contain richer contextual information,which can help identify the authenticity of microblog content.Therefore,this paper proposes a microblog time series division method based on K-Means to divide the microblog events composed of the source microblog and related comments and forwardings.Then,the microblog rumor event detection is completed through the BERT-GRU model,and the detection accuracy reaches 96.7%.Further improving the performance of microblog rumor detection.