Research On Deceptive Review Detection Method Based On Topic Sentiment Model

In the era of the Internet,online review has gradually become the main way for people to communicate and share,and also one of the main channels to understand information.The efficiency and convenience of the Internet greatly accelerate the propagation speed of online reviews.While providing convenience for people,it also leads to the wanton spread of deceptive evaluation information on the Internet,which damages the interests of users and businesses,disrupts the market rules,and even affects the social order.An online review usually exists in the form of text on the Internet,which not only contains the structural information of the text,but also contains the description of the characteristics of things and sentiment expression,that is,the topic and sentiment information of the text.The existing deceptive review detection methods usually use the word frequency,user behavior characteristics,sentiment,structure,and other information of the text to extract the feature vector of the text,then use some classifier to detect deceptive reviews.In this thesis,we will carry out the research on deceptive review detection based on LDA topic model.By learning comprehensively from the advanced ideas of existing related topic models,we explore some improved models of text feature extraction,and then propose new deceptive review detection methods by combine the models with multiple classifiers to achieve better classification effect for true and deceptive reviews.The main contributions of this thesis are as follows:(1)By analyzing current improved models based on LDA,a sentence-level topic sentiment model SJTSM for text feature extraction is proposed,which can simultaneously integrate text topic information,sentiment information and structure information.Then based on the model and by designing a special multi-classifier composed of support vector machine,decision tree and naive bayes classifier,a deceptive review detection method based on the sentence-level topic sentiment model is proposed,which can achieve more effective detection for deceptive reviews.(2)Aiming at the problem that the topic in practical application may be one word or more words(such as noun phrase),and the topic results extracted by many existing topic sentiment models may have ambiguity due to using a single word to represent the topic,a chunk-level topic sentiment model(CTSM)is proposed,which can solve the ambiguity of topic expression to a certain extent.Furthermore,by combining the model with an existing word-level topic sentiment model and the above sentence-level topic sentiment model,a deceptive review detection method based on multi-level topic sentiment models is proposed,which can further improve the recognition ability of deceptive reviews.(3)In order to verify the scientific effectiveness of the above proposed methods,we complete a series of comparative experiments with existing methods on public datasets.The results show that both of the new methods proposed in this thesis have achieved a better classification performance on deceptive reviews detection.