Research On Sentiment Analysis On Product Reviews

In today's online shopping sites,online communities,and social media,text comments have become the most important data source for researchers to study user behavior and understand various phenomena.The rise of shopping diversity on e-commerce sites allows people to buy the goods they need online every day,and at the same time express their feelings and opinions about a certain product at any time.The sentiment analysis of text comments has attracted the attention of researchers in the fields of political science,marketing,communication,social science,and psychology.Analyzing the emotional tendency of review texts,and studying user behavior on online shopping websites and social media is an important research direction.For many years,researchers have been using text reviews to study how users view different products,and perform sentiment analysis on some product reviews on e-commerce websites with different types of topics.From the perspective of comment text,sentiment analysis can be conducted on topics of interest,or sentiment analysis can be conducted on texts such as articles of interest,such as post or Weibo.However,it is of great significance to separate high-quality comments from good and bad comments for sentiment analysis.Compared with traditional machine learning text sentiment analysis methods,although deep neural networks have recently shown significant improvements,the performance of deep neural network models depends on the characteristics of their model design.Information is encoded,and some important information may be lost in some network layers.Moreover,the comprehensive extraction of text features is not too satisfactory,so that the accuracy of text sentiment classification still has room for improvement.Therefore,according to the characteristics of product reviews,how to more comprehensively represent text vectors and improve the quality of text representation features is very important for improving the accuracy of text sentiment classification.In response to the above-mentioned problems,this paper uses deep-level hierarchical concepts on the basis of deep learning technology to build complex concepts on top of simple concepts,and compare and learn directly from raw data.The main research contents are as follows:(1)The representation form of the review text will have a great impact on the sentiment analysis of the text.In order to improve the quality of the review text representation,this paper proposes a text review sentiment analysis model combining CNN and Bi GRU.First of all,traditional convolutional neural networks will not only ignore the contextual semantic information of words,but also lose a lot of feature information during the maximum pooling process,so RNN is used to extract the contextual semantic information of sentences.However,the traditional RNN will cause information memory loss and gradient diffusion,and the use of RNN variants Bi GRU and CNN at the same time can solve this problem well.In CNN,multiple different convolution kernels and maximum pooling operations are used to extract richer and most significant local text features.(2)Although the open source data set has been correctly marked with polarity,its review text contains a lot of redundant and meaningless information.This article rebuilds the high-level data set on the basis of two different open source data sets CHR and CPR.The quality text sentiment analysis data set is used for follow-up work.This paper introduces a suitable attention mechanism based on the model of the fusion of CNN and RNN variant Bi GRU.First,use Word2vec's text vectors to sentences,use these sentence vectors as the input of CNN and Bi GRU,and then fuse the extracted local features with the extracted sequence feature information;finally,in order to improve the quality of the most important words in the sentence Use the attention mechanism to eliminate the influence of some rarely related or irrelevant words.(3)Existing deep neural network sentiment analysis models must use a large number of parameters,need to use complex fitting learning models,introduce comparative learning enhancement mechanisms,and redefine sentiment analysis tasks as a comparison problem,avoiding the use of too complicated The fitting learning mode is obtained,and the similarity score is obtained by comparing the text vector with the labeled sample.Compared with the previous general deep learning models,the model proposed in this paper has achieved better results in accuracy.