Research On Rumor Detection Based On Deep Learning

As people's daily lives become more and more closely related to social media,and Weibo becomes the most important social and information platform.The hidden rumors are influencing netizens' understanding of the incident,and shaking the harmony and stability of the society.For the current social platforms,there are rumors everywhere,but the platforms are too slow to identify and respond in time.Rumor detection methods based on deep learning have attracted more and more attention because of their high efficiency and accuracy.After analysis,the fusion of the sentiment feature of the comment and the textual feature of the main text has positive effect to improve the performance of rumor detection.Therefore,this article proposes a rumor detection method that combines emotional and textual features for Weibo rumors,and makes the following research:Firstly,a method of generating labeled corpus is proposed.Deep learning methods have a common shortcoming,which lacks training data.This method infers the emotional polarity of a comment based on the emotional polarity of the emoji in comment.By using this method,the labeled corpus that could be used for training sentiment analysis model will be greatly expanded,and the performance of sentiment analysis model will be improved,too.Secondly,an attention mechanism that introduces emotional value is proposed.Usually,sentiment analysis is implemented by training a text categorization model and applying it to sentiment analysis tasks.In order to make the model has better categorization performance,this paper chooses to train a text categorization model with an attention mechanism.Besides,in order to make the model more suitable for sentiment analysis tasks,this paper introduces the emotional value of words into the attention mechanism,so that the model training can focus on words which have high emotional value.Compared with other emotion analysis models,the accuracy of the proposed emotion analysis model has been significantly improved.Thirdly,a feature selection method with feature subset selection and redundant feature filtering is proposed.In order to select the textual features of microblog text,for the shortcomings that traditional information gain method is not suitable for the imbalanced corpus,and the selected feature subset has highly correlation degree between classes.This paper proposes a feature selection method which implements the feature selection on imbalanced corpus by limiting the upper limit of the number of features to be selected,and reduces redundant features by deleting features with similar frequency between classes.Using the proposed method and other feature selection methods,text categorization experiments are carried out in KNN classifier,and the categorization results are compared and analyzed.The results show that the proposed method has better performance in accuracy and running time efficiency.Fourthly,a rumor detection method that combines emotional features and textual features is proposed.The sentiment analysis model is used to extract the sentiment features of the detected microblog's comments,and the feature selection method is used to extract textual features of detected microblog's text.The sentiment features and the textual features are combined to express the microblog,and then the rumor detection of the microblog is performed.The method proposed in this paper is used to conduct a contrast experiment of rumor detection with multiple baseline methods,and four kinds of categorization evaluation indicators,precision,accuracy,recall and F1 value,are analyzed and compared.The results show that the method proposed in this paper has better performance than other baseline methods.