Research On Personalized Recommendation Based On Emotional Analysis Of Product Reviews

With the rapid popularization of the Internet,more and more people express their ideas and suggestions on the Internet.These data contain people’s emotional information about something and also imply people’s preferences.However,the consequent problem of information overload constantly interferes with people’s life,and it becomes very difficult for people to find valuable content from a large amount of information data.Through the research of scholars,the recommendation system has effectively solved this problem,and has been widely used in various fields.In the field of e-commerce,recommendation systems are used to analyze product comment information,uncover user interests,and provide personalized recommendations to each user.This not only improves the user’s shopping experience,but also allows merchants to discover their own shortcomings and optimize their services.However,the existing recommendation methods still have some shortcomings:firstly,the expression of the word vector is not comprehensive enough when the vectorization of the word is carried out in the commodity review text;Secondly,the user’s characteristic information cannot be fully learned in the semantic analysis of product review text.Thirdly,it is not enough to explore the user characteristics and product characteristics,which is easy to cause the inaccurate recommendation.On this basis,this thesis combines thematic features and semantic features to obtain the word vector of product reviews.Deep learning technology is used in sentiment analysis and recommendation tasks.The main research contents are as follows:(1)In terms of word vector expression,after denoising and word segmentation of acquired data,the LDA model is used to obtain a word vector containing the topic features of product reviews.The BERT model is used to integrate the features of word information and location information into the word vector expression,and word vector containing the semantic features of product reviews is obtained.By combining the word vectors of the two,a word vector matrix integrating the thematic features and semantic features can be obtained,which can more comprehensively express the feature information of product reviews.(2)In the sentiment analysis module,an emotion classification model of LDA-BERT-CNN-Bi LSTM(LBCB)is proposed.Firstly,CNN is used to obtain local features of product reviews,and Bi LSTM is used to obtain contextual semantic features of product reviews.This combination neural network model can make up for the deficiency of single network model CNN or Bi LSTM in extracting feature information.Moreover,Bi LSTM uses two LSTM networks with opposite sequences.To better capture the context of the text,get the above and below information of the comment separately.Secondly,the two features are integrated to obtain the local feature information and contextual semantic features of the text,making the product features acquired more comprehensive.Finally,the attention mechanism is introduced to learn important words in product reviews and give higher weight to important words,so that the model pays high attention to the information of key words.In this paper,the effectiveness of the LBCB model is verified by comparative experiments.The experimental results show that the proposed model can improve the accuracy of classification results.(3)In the personalized recommendation module,a personalized recommendation model of LBCB-Bi GRU(LBCB-B)is proposed.Firstly,the vectorized representation of the obtained product review text is carried out.Secondly,model user reviews and product reviews separately,and extract user feature information and product feature information respectively through the LBCB sentiment analysis model above.Finally,user characteristics and product characteristics are integrated.Since the interactive relationship between users and commodities is developed based on time series,Bi GRU learns the nonlinear relationship between users and commodities to obtain the predicted score of users’ treatment of recommended commodities.The experimental results show that compared with the traditional recommendation model,the recommendation model proposed in this paper can improve the accuracy of prediction results.