Research On Deceptive Reviews Detection For Multiple Domains

With the development of e-commerce,more and more individuals and business organizations use online reviews to make purchasing decisions.Positive reviews can render significant financial gains and fame for businesses,which provides a strong impetus for deceptive reviews.In the past few years,the problem of deceptive reviews has become widespread.The detection of deceptive reviews is an urgent and important topic,which is crucial to ensure the credibility of information on online platforms.Based on this,the following two tasks were carries out:Concerning the problem that traditional discrete models fail to capture global semantic information over a sentence in deceptive reviews detection,a hierarchical neural network model with attention mechanism was proposed.Firstly,different networks were adopted to model reviews,to explore which one could get best semantic representations.Then,reviews were modeled from perspectives of user and product,in which the former focused on the user's preferences and the latter focused on the feature of the product.Finally,the two representations learned were combined as the final sentence representations.The experiments were done on Yelp dataset,using the accuracy score as the evaluation indicator.Results show that the proposed hierarchical neural network model with attention mechanism has achieved the state of the art performance.The accuracy scores are higher than traditional discrete methods and previous neural baseline systems by 1.0 to 4.0 percentage points.Thus,the proposed neural network achieves better results.Concerning the problem of poor performance of detection deceptive reviews on cross-domain,a detection of deceptive reviews model based on adversarial training is proposed.Firstly,based on the mixed data sets of the three domains,different machine learning models are used to model the reviews to verify the effectiveness of the proposed model.Secondly,based on the data sets of each domain,the deceptive reviews detection is carried out.That is training model on the data of one domain to verify the classification performance on the other two domains.Specifically,small perturbations were made to the input and LSTM model was used for adversarial training,then a final prediction was output.The results show that the proposed model has better generalization ability than the previous model.