Research And Application Of Fine-grained Sentiment Analysis Of Online Reviews For Personalized Recommendation

As the information technology is being well-rounded and mobile network is being universally popular,meanwhile,the rapid development of E-commerce platform triggers relatively serious problem of information overload.As a sharp weapon to alleviate the problem of “information overload”,personalized recommendation has been paid more and more attention.However,the key problem is that the traditional recommendation system based on the scoring data can not realize the accurate user interest mining.And the result is lack of interpretation.As the creator of information,people are active in all kinds of online platforms,with the result that they leave massive interacting data.Very rich information of users' aspect-based sentiment orientation is contained on the data of user generated comments.And this kind of information can realize accurate user interest mining and reflect all respects of characteristic information of the product,which is massively suitable to apply to some tasks,such as personalized recommendation,product feedback,etc.However,most of such product reviews are long texts,it is universally that there are some problems of text information redundancy and sparse features with long text.In view of the above problems,this essay construct a model of Fine-grained Sentiment Analysis,which has stronger performance,to identify the sentiment orientation of specific aspect of the text effectively,then apply the sentiment orientation information to personalized recommendation.The works are as follows.(1)This paper builds a fusion model for Fine-grained Sentiment Analysis called Bert-Bi GRU-CNN-Att.The model use Bert(Bidirectional Encoder Representations from Transformers)to replace the traditional word embedding layer of model,achieve the full mining of context-related text semantic information and alleviate the problem of sparse feature distribution.In addition,this model combines the advantages of Bi GRU for processing time series data and the powerful local feature extraction capabilities of CNN and Attention mechanisms to alleviate the problem of information redundancy.Experimental results prove that the performance of this fusion model on the tasks of Fine-grained Sentiment Analysis is better than other comparison models.(2)Based on the idea of ensemble learning,this paper proposes an integrated model called Union-Bert,which further improves the performance of Fine-grained Sentiment Analysis model.The main approach is that this paper use different pre-trained language models(ERNIE ?Roberta?Bert-wwm-ext)to replace Bert and build multiple models for sentiment analysis.Then,this paper used the aspect-level sentiment annotation data set and sentence-level sentiment annotation data set of the restaurant to supervised train these models,and finally obtained the integrated optimal integrated model Union-Bert based on the integrated algorithm.The experimental results show that the performance of the integrated model in this paper is improved compared with the single model.(3)This paper proposes a hybrid recommendation model based on Union-Bert and NMF.Firstly,based on Union-Bert,this paper proposes a calculation method of Aspect-Based Sentiment Recommendation Index,which is explainable.And a algorithm is designed to use the recommendation index to achieve the task of score prediction.Secondly,this paper combines the recommendation index with the real score to realize the hybrid recommendation based on Union-Bert and NMF.Finally,this paper designs several groups of hybrid recommendation models for experimental comparison.The experimental results prove that the performance of the hybrid recommendation model in this paper is better than other comparative models in score prediction task.