Sentiment Analysis Based On Deep Learning

E-Commerce platform,as an important vehicle for digital consumption,has opened up a new sales channel for some brands with distinctive features,unique stories and consumer cognition in the process of supporting and promoting local cosmetic brands.The consumption model that uses reviews as a reference is gradually expanding.Consumers could make their decisions based on reviews of the product.Reviews could also help merchants understand consumers’ sentiment and help companies conduct product optimization,customer profiling and market decisions,thus increasing their economic benefits.In the face of fast-growing data,if we develop tools that could automatically analyze the sentiment of product reviews,uncover valuable item attributes and emotional orientations from a large number of reviews,we could enable customers to choose their target products better,which could also help understand customers’ perceptions.This thesis takes Tmall domestic makeup products as the research object,and collects review text data of representative brands for sentiment analysis.However,the simple sentiment classification is not enough to fully explore the value behind the review data,and the research on the sentiment analysis of the review text of the products needs to be further deepened.So we explore both coarse-grained and fine-grained aspects with respect to users’ sentiment(1)Coarse-grained is mainly to determine whether the overall sentiment tendency of the reviews is biased towards positive or negative.Firstly,deep learning algorithms based on Word2 Vec combined with CNN and Bi LSTM model is used to train the corpus package which has been labeled with sentiment attitude.Then the review texts of national makeup products are put into the trained classifier to determine the overall sentiment tendency;(2)Fine-grained refers to mining the user’s evaluation features from the perspective of various aspects of product.The fine-grained sentiment analysis is studied from two perspectives: sentiment analysis study of all brand products and the favorable ratings for each attribute of different brands.In the first part,Word-Cloud and LDA topic model are used for visualization and topic analysis respectively.This part presents attributes sentiment analysis for product and review sentiment score for brands in visual form,to achieve a comparative analysis of the emotional tendencies of all products.The second part combines the perspective of brand marketing strategy,and the products of different brands are analyzed from each attribute in a horizontal comparison to explore the strengths and weaknesses of each brand in each attribute.The experiment shows that(1)The accuracy of the original CNN model and FCNN model are low,while the improved Word2Vec-CNN model and Word2Vec-Bi LSTM model have a large improvement,which proves the effectiveness of the Word2 Vec mode,and the Word2Vec-CNN model performs better than the Word2Vec-Bi LSTM model;(2)For all products,the degree of consumer sentiment could be visualized.For different brands,the strengths and weaknesses of each brand could be clarified,which could help consumers understand the products more comprehensively and specifically,while merchants could optimize their products and bring more economic benefits.