Sentiment Analysis Of Bi-LSTM Web Review Text Integrating Sentiment Dictionary And Attention Mechanism

In the era of the rapid development of the Internet and the explosion of data,text has become the most common way of expressing people’s emotions.Sentiment analysis of the text can not only provide reference for consumers,but also improve the overall service mechanism of merchants.Therefore,sentiment analysis of the text has an important research significance.This paper adopts the method of combining sentiment dictionary and deep learning to conduct sentiment analysis on the food and beverage review data crawled from a review website.The main research contents and experimental results of this paper are as follows:(1)After acquiring and preprocessing the data,this paper constructs the basic sentiment dictionary of online review texts.For the basic sentiment dictionary of online review texts,this paper selects the results of de-duplication and merging of the three dictionaries of HOWNET,TSING,and NTUSD.A total of 8,243 positive sentiment words and 12,238 negative sentiment words are obtained.For the expansion of emotional words in the professional field,this paper proposes the dictionary expansion method based on point mutual information and correlation coefficient.This method improves the accuracy of dictionary expansion by introducing correlation coefficients on the original point mutual information model.Then the extended model proposed in this paper is compared with the model using the correlation coefficient alone and the point mutual information alone.The results show that the model with the correlation coefficient introduced on the basis of the original point mutual information model has the best training effect,the point mutual information is the second,and the correlation coefficient is the worst.(2)The sentiment analysis is carried out on the online review text.The sentiment analysis model of this paper is improved on the basis of Bi-LSTM,and the sentiment analysis method Att-Bi-LSTM based on attention mechanism and Bi-LSTM is proposed.This method enhances the extraction of feature vectors and improves the performance of the model by introducing the attention mechanism,the linearly transforms of the extracted feature vectors through the full connection method,and then introduces them into the Softmax classifier to classify the sentiment of online review texts.Comparing the experimental results of the Att-Bi-LSTM model with the CNN model and the Bi-LSTM model,the results show that the score of the Att-Bi-LSTM model is 1.87% higher than that of the Bi-LSTM model,which is higher than that of the CNN model.2.44%.(3)This paper continues to improve the Att-Bi-LSTM model,and proposes a sentiment analysis method Dic-Att-Bi-LSTM based on the sentiment dictionary and the Att-Bi-LSTM model.This method enhances the model’s attention to the semantics in the online review text and improves the accuracy of the model by adding a sentiment dictionary.Comparing the Dic-Att-Bi-LSTM model with the Text CNN model and the Att-Bi-LSTM model,the results show that the Dic-Att-Bi-LSTM model has a higher accuracy rate,recall rate or score.Outperforms Att-Bi-LSTM model and Text CNN model.