Research Of Sentiment Analysis Based On Product Reviews

With the popularization of the Internet and the rapid development of information technology,a large number of subjective comments appear on weibo taobao,Tmall,and other websites.These comments contain the sentiment information and subjective opinions on the evaluation object.Users are used to obtaining valuable information from various comments to assist their decision-making.Sentiment analysis is a comprehensive research subject involving artificial intelligence,natural language processing and machine learning.At present,researchers of sentiment analysis about text research methods based on sentiment lexicon and machine learning,but most of the traditional methods of sentiment analysis require a lot of manual work,and they cannot get good performance.The sentiment analysis studied in this paper is binary(positive and negative).In order to improve the sentiment classification effect of product review text,this paper set up several groups of comparative experiments to select the optimal classification model and mainly did the following work:(1)by exploring new sentiment words to expand the sentiment lexicon,the coverage of the sentiment lexicon is improved.It includes two methods: sentiment lexicon expansion based on synonyms and rule templates.The experimental results show that the extended dictionary can effectively improve the classification effect.(2)The sentiment analysis based on machine learning is studied.In the shallow learning,the traditional feature selection methods such as unary phrase and binary phrase are changed,but the word vector is taken as the feature input,and the sentiment information and polarity transfer are integrated into it,so that the generated classifier can obtain deeper semantic information.The feature representation method avoids the problems of ignoring semantics,polarity transfer and high feature dimension in traditional methods.In deep learning,considering the advantages of long and short time memory networks,the bidirectional LSTM model based on the attention mechanism was selected as the classification model of deep learning.Experiments show that deep learning algorithm is better than shallow classification algorithm in sentiment classification.(3)The sentiment analysis method based on the word vector technology is studied.We use FastText model and BERT model respectively for the study of sentiment analysis study.FastText is similar to word2 vec principle.BERT seems to be a word2 vec enhanced.Their level of word or sentence level vectors are pre training.The vector representation of word2 vec is context-free,while BERT is contextual.The results of this set of comparison experiments show that BERT performs better in classification.(4)The sentiment analysis method of fusion technique is studied andachieved.Mainly sentiment classification of text based on Bagging algorithm which uses multiple weak classifier to jointly determine the classification results.At the same time,according to the current development situation of classification model,the combination of the sentiment lexicon and the BILSTM model based on the attention mechanism is adopted to conduct a comparative experiment of Bagging algorithm.The experimental results show that Bagging algorithm has a relatively higher classification accuracy.