Research On Topic-Sentiment Joint Probabilistic Model For Detecting Deceptive Reviews

With the development of intelligent terminal devices and popularity of e-commerce,consumers tend to shop online which bring a boom in e-commerce.Due to the problem of information asymmetry in the virtual market,consumers could not accurately obtain quality of goods so they learn peripheral vision of what the quality of the goods or services with the help of historical reviews.However,for the purpose of improving their reputations,some illegal businesses exploit the vulnerability of mechanism and employ online ghostwriters to fabricate fake reviews which entices consumers to buy.Moreover,the generation of big data makes it difficult to identify deceptive reviews by the naked eye.Though,traditional methods of detecting deceptive reviews make some progress,it restricted improvement of accuracy.Because these methods almost utilize language features and syntactic features which cannot mine semantic and sentiment information very well.Moreover,deep neural network models can get abstract representation of text but lack of interpretability.To improve the accuracy of detecting deceptive reviews,this dissertation propose an Unsupervised Topic-Sentiment Joint Probabilistic Model(UTSJ)from the view of semantics and sentiments.The work of this dissertation 1s outlined as follows:1)Integrate topic models to detect deceptive reviews.From the view of semantics and sentiments,this dissertation apply tradition topic model LDA and topic model integrated with sentiment JST on field of detecting deceptive review and verify the performance.Based on traditional topic model,this dissertation proposed Unsupervised Topic-Sentiment Joint Probabilistic Model(UTSJ)and compare with the baseline models.2)Implement several groups of comparative experiments on real-life yelp reviews.First of all,restore traditional detection models of deceptive review.Select the models based on text language feature and model based on shallow syntactic feature as baseline.Secondary,respectively construct balanced dataset and unbalanced dataset on two domain of yelp reviews(hotel and restaurant)and investigate the detection performance of these above models under such conditions.Experimental results show that the model proposed in this dissertation is superior to baseline models under the condition of balanced dataset and unbalanced dataset and more suited to apply in e-commerce environment(unbalanced large sample).