Research On Review Sentiment Analysis With User And Product Information

Review sentiment analysis with user and product information is an important task in sentiment analysis,which aims to infer the sentiment score(also called sentiment classification)of review documents according to three kinds of information: a review document,the user(the user who makes the comment)and the product(the product targeted by the comment).Recently,sentiment classification models based on deep learning have achieved significant performance compared with traditional methods,which is based on artificial features by automatically learning distributed features,and have become one of the hot topics in current research.Although the review sentiment analysis model based on deep learning with user and product information has achieved good results,it still has deficiencies in the following two aspects: 1)most of models used the same way(for example,the attention mechanism)integrated user and the product information,not fully consider the user and the product information on the different effects of sentiment score;2)The semantic vector representations of users and products are mostly learned implicitly,and the relationships between users and reviews,products and reviews,similar users and similar products are not explicitly modeled.Focusing on the above problems,this paper carries out research on review sentiment analysis with user and product information to further improve the performance of sentiment analysis,mainly including the following two research contents:1)Knowledge Mutual Distillation Based Review Sentiment Analysis Method.In view of the first deficiency,it is proposed to integrate user and product information in different ways.Specifically,the semantic fusion layer is used to integrate user information,and the attention mechanism is used to integrate product information.Further,a training framework based on knowledge mutual distillation is proposed.In the training framework,The main model that integrates user and product information and the auxiliary model that integrates user and product information are jointly trained,and knowledge is transferred through the method of knowledge distillation to enhance each other,so as to make full use of the different influences of user and product information.The experimental results show that the proposed method achieves better results than similar benchmark methods on review sentiment classification.2)A review sentiment analysis method based on hierarchical transformers and graph networks.In view of the second deficiency,firstly,the initial semantic vector representation of users,products and documents is learned based on the hierarchical Transformer model;then,a graph neural network is proposed to explicitly model the dependency between users and review documents,between products and review documents,dependency between similar users,similar products,and similar review documents to better learn semantic vector representations of users,products,and review documents.Experimental results show that the proposed method outperforms the baseline methods on commonly used datasets.