Research On Sentiment Classification Of Product Reviews Based On Recurrent Neural Network And LDA Model

With the rapid development of the Internet,online shopping has become an inseparable part of people's daily lives,and the number of Internet product reviews also grows exponentially.Mining the emotional tendency in these reviews not only can help consumers determine the purchase intentions quickly,but also can help sellers understand the degree of satisfaction of products for consumers,analyze strengths and weaknesses of the product,and make sales decisions.However,with such huge amount of product reviews,it's time-consuming and subjective to get sentiment tendencies of reviews based on manual analysis.Therefore,it's significant to study how to analyze the emotion automatically for the mass product reviews?This thesis mainly studies the emotion classification of product reviews from two aspects:the coarse-grained emotion classification of the product reviews and the fine-grained property emotion classification of the product reviews.The specific innovative achievements are as follows:(1)A product reviews emotion classification method based on LSTM for bidirectional GRU is proposed.Machine learning method that used for emotion classification is based on shallow features,and Classifier performance is limited.In this thesis,Word2 vec is used to construct deep learning features.LSTM model based on recurrent neural network is used as the classifier model,which can store the information in front of the sequence and solve the problem of long-term dependence and gradient explosion.Then,aiming at the problem that the model can only capture the above feature information,a bidirectional GRU model which can capture the context feature information is proposed.The experiment results show that the LSTM model achieves a 90.03%accuracy rate based on the model itself,is 8.9% higher compared with the best performing SVM model in machine learning.Using bidirectional GRU algorithm promotes the accuracy from90.03% to 92.85%.(2)A product reviews emotion classification method based on large-scale sentiment lexicon and bidirectional GRU is proposed.The bidirectional GRU model needs to be annotated manually,so it has the domain dependency,individual subjectivity and cost waste of manpower.An improved algorithm based on large-scale sentiment lexicon and bidirectional GRU is proposed.Experiment results show that the improved algorithm combining the two algorithms achieves an accuracy of 93.96% and is 1.11% higher than the bidirectional GRU model,which isan increase of 5.33% compared with the algorithm that combines emotion dictionary and SVM.(3)A method of attribute and emotion word extraction based on SC-LDA model is proposed.Aiming at the defect that some improved theme models have low rate of extracting low-frequency synonym attribute words and sentiment words,this thesis introduces semantic constraints into standard LDA model,and proposes SC-LDA model to improve the recognition of theme words,the degree of discrimination and the extraction rate of their relationship.Experiment results show that the model improves the accuracy to 83.61%,which is 12.67%higher than that of ASUM model.(4)An emotion classification method based on SCEB-LDA model is proposed.At present,some improved theme models appear low cohesiveness and distinctiveness of thematic attributes,the semantic relevance between attribute and emotion words is not high enough,the semantic understanding of the topic is not clear.In this thesis,the semantic constraints and the affective allocation constraints are introduced into the standard LDA model,then a SCEB-LDA model is proposed.The experiment results show that the model improves the accuracy to 82.6%,which is3.66% higher than that of HASM model.