Research On Deceptive Opinion Recognition Based On Deep Learning

With the development of the Internet and mobile terminals,electronic commerce has become an integral part of people's daily life,followed by the rapid growth of commodity information and user reviews.User reviews play an important role in electronic commerce,because users use comments as reference resources to measure the quality of the commodity in online shopping,and will influence the decisions of consumers.Therefore,the online sellers hired people to give their own products fraudulent positive comments and abuse competitors by writing negative ones for more profits,which brings serious effects to the healthy development of electronic commerce.Previous studies showing that the accuracy of distinguishing such comments by common people is low.In order to effectively identify such comments,many researchers have achieved certain results by using shallow semantic features.On the task of identifying deceptive comments,this paper takes deep learning approaches as the focus of research because it can dig deeper semantic features.The research contents are summarized as follows:(1)Deceptive opinion recognition based on traditional model.Four classifiers are adopted in traditional model.According to the characteristics of the deceptive comments,we generated four kinds of features(text-related feature,emotional feature,psychological feature and syntactic-related feature,respectively).We proposed a voting mechanism which combined the results of several models,which perform better than the baseline method.(2)Extending corpus by semi-supervised learning.In order to solve the data scarcity problem of deceptive comments,we use a semi-supervised learning algorithm to extract comments,which has a high confidence level,from network resources crawled by Web crawler,and then add the comments to corpus.(3)Deceptive opinion recognition based on deep learning model.In the task of deceptive comment recognition,using word embedding as input,we conduct experiments on LSTM,CNN and the fusion of the above models.The experiment results show that the hybrid model of CNN and LSTM is best,and the accuracy rate is 2 percentage points higher than that of baseline method.(4)Deceptive opinion recognition based on Attention mechanism.This paper implements two attention mechanisms,namely,the feed-forward attention model and the context-based attention model.The semantic encoding obtained by attention model can explain sentences better because it highlight keywords of sentences.In this paper,the attention mechanism is applied to the LSTM model and a hybrid model of LSTM and CNN,which further improves the accuracy.