Recommendation Model Research Based On Reviews Text Analysis

As the exponential growth of Internet scale,information overload problem is increasingly serious to internet users.Recommendation system for different users with personalized recommendation,is seen as a promising way to solve the problem of information overload.But traditional recommendation methods,such as collaborative filtering,only considered the users' ratings of commodity(implicitly or explicitly),and ignored the many user reviews which are available.The purpose of our research is to extract structured and meaningful information from the user's comments on the goods by text mining and deep learning technology,and combine such information to the traditional recommendation model,in order to solve the shortcomings existed in traditional recommendation methods,and to enhance the robustness and reliability of the recommendation system,also improve the recommendation results.This paper first introduces typical research achievements of scholars at home and abroad with in these three areas: reviews text analysis technology,recommendation system,and reviews text based recommendation system.And we analyzed the problems and advantages of the traditional methods,in order to improve their performance and combine them together.First,in the process of analysis of the reviews text,we take use of the feature level opinion mining,sentiment classification and doc2 vec language model which based on deep neural network to get the implicit representation of each review.Second,in the process of analysis of the score matrix,the hidden factor model based on matrix decomposition is used to obtain implicit representation of the user and goods.Finally,using the method based on probability graph and the method based on multi-task learning,an innovative recommendation model is proposed,which seamlessly combines the doc2 vec model and the classic matrix decomposition based recommendation algorithm.The comparative experiments on different sets of algorithms on Chinese and English data shown that the proposed model could effectively make use of the reviews implied semantic information and improve the score of rating prediction tasks.