Research On Fake Review Detection Based On Topic And Language Clue

With the rapid development of the Internet,online reviews have become an important channel for people to exchange and share consumption and shopping experience.The quality of reviews has a significant impact on consumers’ purchase decisions,which indirectly affects the profits of businesses.Therefore,in order to protect their own interests,some businesses have taken the form of "paid order brushing","positive feedback cash back" and other forms of mass production of fake positive reviews,resulting in today’s e-commerce platforms being flooded with fake reviews.Fake reviews have a great negative impact on consumers,businesses and platforms.First of all,fake reviews will mislead consumers’ judgment which leads to purchase something improper.Secondly,the value of reviews becomes lower.It is difficult for businesses to gain consumers’ demands and suggestions in a timely manner.Finally,the platform will also suffer from reputation damage and customer chum due to too many fake reviews.However,the detection of fake reviews is not easy.Because of the concealment and complexity of fake reviews,manual detection becomes rather difficult.What are the characteristics that distinguish true reviews from fake reviews? How to detect fake reviews automatically? These two issues are the mainstream of current research.In response to the first question,we conduct LDA topic modeling and linguistic clue analysis for true reviews and fake reviews respectively.Different from the previous LDA model based on Unigram,the LDA used here is based on Bigram.Compared with Unigram,Bigram-LDA has better interpretability.In addition,we build the text co-occurrence network to check the distribution differences of the subject words of true reviews and fake reviews when analyzing the differences of the topic words.Finally,we use LIWC to analyze whether there are differences between true and fake reviews in the four linguistic clues of emotion,cognition,social interaction and perception.The results show that compared with fake reviews,true reviews contain more detailed information,such as special food and special activities.Besides we have found that a single true review may contain quite rich topic information,while fake reviews are few.In addition,we find that there are significant differences between real comments and false comments in the four categories of linguistic clues.In response to the second problem,we propose a fake review detection model that combines topic model,linguistic clues and deep neural network.The model consists of two modules,one is the Attention based Convolutional Neural Network model that mechanism to extract implicit semantic features from text information.The other is the Bi GRU,which extracts important features from the topics and linguistic clues.Then we integrate the two modules to detective fake reviews.When compared with other baseline models,this model exhibits outstanding performance,with an improvement of about 5%.Past research has provided rich features and models for fake review detection.Based on previous research results,we creatively proposes to use the Bigram LDA topic model and four linguistic cues of emotion,cognition,social interaction and perception to extract the features of comments.And then we provide reasonable suggestions on how to identify fake reviews from the aspect of topic words and linguistic clues.In addition,the two kinds of features are fused with the deep neural network to improve the classification effect of the model.The results of this paper enrich the theoretical research of topic,linguistic clue analysis and fake review detection.In addition,we also provide a reference scheme for the platform to build an effective fake review detection model.