| With the development of the network environment and culture,Internet users are keen to share and publish online reviews on social media platforms or review websites,and freely express their views and opinions on hotels,restaurants,movies,e-commerce products,mobile applications,etc.for other users to browse.However,due to the current development and prevalence of cultural diversity,or driven by the interests behind,unfair and untrue online reviews(also known as spam reviews)frequently appear and penetrate into various fields.The proliferation of spam reviews disturbs consumers’ judgment and is not conducive to the healthy development of market competition.In order to identify online spam reviews and malicious users who post spam reviews,researchers have designed automated detection methods based on machine learning and deep learning.Douban,a typical domestic movie review platform,has been attacked by spam reviews in recent years.Therefore,this study takes movies on Douban as the research object.The main research work and innovations are as follows:1.This thesis first collects real Douban online reviews and related movies information,and through in-depth analysis of the data and the background of the reviews,four categories of Douban reviews are established,namely spam positive reviews,spam negative reviews,true positive reviews and true negative reviews,this classification is more detailed than the binary classification proposed in previous studies.This thesis proposes reliable and effective labeling rules for Douban data,and constructs a reliable movie review dataset.Through the data mining of reviews in various categories,it is found that Douban reviews have three characteristics:organized action,antagonistic groups,and burst attack.2.This thesis proposes a co-attention based feature fusion network for spam review detection,which fully extracts effective features for review data,reviewed works and review senders.The comparison experiments with the baseline models show that the model achieves the best classification results on both the Douban review dataset and the English review dataset.The effectiveness of each module of the model and the effectiveness of attention weights are respectively proved by ablation experiments and case analysis.3.In order to improve the robustness and generalization ability of the neural network model to resist the attack of data noise,this thesis proposes a co-attention based deep model with domain-adversarial training for spam review detection.This thesis generates noisy data based on the original review data,and simulates the attack of the noise data on the model.Through comparative experiments,it is proved that the model with domain-adversarial training can effectively improve the robustness of the spam review detection model. |
