Research On Lightweight And Interpretable Rumor Detection Methods

With the rapid development of Web2.0 and mobile Internet,it provides people with a convenient and efficient information exchange platform,but it also reduces the cost of spreading rumors and increases the speed of spreading rumors.The demand for automated rumors detection algorithms is becoming more and more urgent.The traditional rumors are organized according to time series,and the detection algorithm that uses the deep learning model to encode rumors loses the structural information of the rumors.Recent work proposes to comprehensively encode the text and structure of rumors based on deep models such as recurrent neural networks or graph convolutional networks,which effectively improves the performance of rumors detection algorithms.However,the current rumor detection algorithms based on deep learning models have the following shortcomings:(1)The existing rumor detection models based on deep models have complex structure,low efficiency,and are not lightweight enough to meet the requirements of large-scale rumor detection.(2)The rumor detection algorithm based on the deep learning models have made a huge breakthrough in performance,but it cannot provide a theoretical explanation for the performance improvement,and the algorithm lacks interpretability.In view of the above problems,the main research work of this paper is as follows:(1)This paper proposes a rumor detection algorithm PPA that encodes rumor text based on rumor propagation path and enhances the rumor representation based on path representation.This method first constructs a rumor propagation structure tree based on the dataset,and then extracts the rumor propagation path based on the rumor propagation tree,using the propagation path as the basic coding unit instead of tree nodes;secondly,the PPA model filters paths based on the attributes of the propagation path,and uses lightweight model to encode propagation paths;finally,based on the representation of the rumor propagation path,the representation of the rumor propagation structure tree is obtained based on the maximum pooling and used for classification.Experiments on three public datasets show that the model PPA proposed in the paper has achieved excellent performance comparable to the existing complex models,but is more lightweight.(2)This paper proposes a rumor detection algorithm PPA-WAE based on neural topic model to enhance the representation of rumor propagation path.The model first builds an unlabeled response path text dataset based on the rumor propagation structure tree;then introduces a neural topic model,pre-trains the neural topic model based on the response path text dataset,and analyzes the distribution of emotional topics contained in the response path;after the neural topic model is pre-trained,the encoder of the neural topic model is used to obtain the representation of the response path,and then the PPA model is jointly trained to further improve the performance of rumor detection.Experiments on four public datasets show that the PPA-WAE model proposed by this paper is better than the existing complex models in performance,and can explain the improvement of rumor detection performance from the perspective of the text topic,which improves rumor detection interpretability of the algorithm.The source code and data have been published in the https://github.com/zperfet/Path Fake.