Linguistic Feature Based Automatic Detection Of Deceptive Opinion

With the fast development of internet technology, internet has becomeone of the most important paths for people to achieve information. Thisinformation makes great effect on people’s decision every day, especiallypurchasing decision. Driven by these tremendous profits, a group ofprofessional writers who earn money by sending opinion spam begin tospread spam information and deceptive opinions. The opinions theypublish in online forums have seriously affected the normal users’judgments on value of goods. These non-objective opinions mixed innormal opinions makes it more difficult for netizen to achieve theinformation they need.In this thesis, our standard dataset is collected by simulating theemployers of “Water Army”. Based on this dataset, we try to use featureextraction methods and classification models to classify these opinions.Based on this framework, genre identification method, psycholinguisticdeception detection and text categorization method are used to extractfeatures. Based on the above features, Na ve Bayesian (NB), SupportVector Machine (SVM), Maximum Entropy Model (MEM)，ArtificialNeural Network (ANN) and multi-state neural network (MNN) are testedto compare the performance of the classification algorithm in our task. Wealso apply this framework on Chinese normal spam filtering task andachieves good performance as well.Based on the basic framework, we also achieve further improvementon both aspects of feature extraction methods and classification methods.For feature extraction part, deep linguistic features are proposed torepresent the linguistic habit of people and achieve better performance than basic feature sets. For classification method part, multi-state neuralnetwork is proposed and applied to this task. This method achieves betterperformance than traditional artificial neural network to some extent.