| With the popularity of the Internet and social media,fake information spreads faster and more widely,and the number of it keeps increasing.Traditional manual feature extraction and classification methods are difficult to meet the needs of large-scale fake information detection,while deep learning has powerful capabilities of the feature extraction and classification,which can effectively process large-scale data.Fake information has various forms,including text,images,videos and other forms.Deep learning can learn feature representations of multimodal data better by processing different forms of fake information,and deep learning can learn higher-level abstract features,which can better distinguish true information from fake information,improve the accuracy and robustness of fake information detection,and better cope with current challenges of fake information detection.The main contributions of this paper are as follows:(1)In order to make full use of text sentiment features for fake information detection,this paper proposes a fake information detection model based on the BERT sentiment analysis.Based on BERT model,this model adds a text sentiment polarity prediction task to make BERT model learn word sentiment polarity in different contexts.And it uses the word sentiment polarity as an auxiliary feature input to subsequent Attention-Bi LSTM model for fake information classification.Experiments are conducted on the Weibo dataset,and the accuracy of the model proposed in this chapter is generally higher than that of other models.(2)Considering that there is also rich visual information in images,this paper proposes a new fake information detection model based on attention and multimodal fusion mechanism.This model firstly uses BERT to extract text features and adds text sentiment prediction task proposed in Chapter 3 to BERT to enhance text sentiment feature learning;then it uses VGG-19 to extract image features;then it inputs text and image features into the self-attention layer;finally,it establishes attention layer through cross-modal encoder to realize text-image feature fusion.And it inputs fused information into fake information detector for classification.The Weibo and Twitter datasets are tested on the models.The experimental results demonstrate that the proposed models have higher prediction accuracy than other models.(3)This paper designs and implements a new system for detecting fake information.The system has following characteristics: First,it provides a front-end interface where users can input texts and images,and view prediction results;Second,it uses fourth chapter proposed attentionbased and multimodal fusion mechanism for detecting fake information as back-end,which can make full use of multimodal data for fake information detection;Then,it visually displays prediction results to users.This paper conducts functional testing on system.Results show that the system can effectively detect fake information and provide users a good experience. |
