Study On Network Rumor Detection System Based On Deep Reinforcement Learning

With the advancement of technology,social media accounts for a growing proportion of people's lives,and has become the main way for people to grasp news and information.There is a huge amount of unverified and even contradictory information on social media.The spread of rumors will mislead people,affect people's lives and become an uncertain factor in society.How to recognize the news on social media and reduce the bad influence caused by rumors is an important issue.Rumor detection is a text classification problem,which is to determine whether a post is a rumor by extracting text features of the original posts and replies,user features,or posts spread features.In this paper,we propose a new rumor detection model.The model contains interpretability and early discovery features.The principal contents are as follows:1)Previous rumor detection models lacked interpretability.In this paper we propose a rumor detection model based on hierarchical attention mechanism.Since sentences and words have different importance for this task,this method combines word-level and sentence-level attention mechanisms.This model calculates the weight of different words in the sentence and the relation between different comments and the original post through the two-layer Bi-GRU and attention mechanism,and finds out the most valuable words and sentences.2)Rumors are very real-time.The model does not have a high application value,if a rumor can only be judged when it has been spread in a wide range and have a great impact.Given the problem that most current rumor detection models only consider detection accuracy but do not consider early rumor detection,this paper proposes an early rumor detection model based on DQN.The experiment shows that compared with the single rumor detection model,the model combined with DQN has higher accuracy in the early stage of rumor generation.3)This paper designs a rumor detection prototype system.The system includes a message crawling module,data preprocessing module,algorithm detection module,and user interaction module.The message crawling module will crawl the recent rumor,and provide data for the system's rumor display and query function.The algorithm detection module provides rumor detection function.The data preprocessing module changes the primitive data into a pattern that can be used by the algorithm.