The Research Of Fine-grained Sentiment Analysis For Product Comments

Under the web2.0 era background,a large number of emotional texts have appeared in the network.Product comment is one kind of the emotional texts,which is some real feelings that consumer shared about product in the online shopping platform.With the development of e-commerce,product comments increased rapidly.Analyzing these product comments can provide more decision basis for merchants and consumers.Traditional sentiment analysis is coarse-grained sentiment analysis that oriented on chapters and sentences,which cannot meet the requirement that consumers want to know more about the detail of products.This paper analyzes product comments by fine-grained sentiment analysis method,which contains the recognition of sentiment elements and appraisal expression.The detailed work is as following:Firstly,the traditional sentiment analysis ignores comment target.This paper proposes a fine-grained sentiment analysis method based on CRF to extract sentiment elements.To extract sentimental information feature,we expand the basic sentiment lexicon.To improve the efficiency of corpus annotation,we propose tri-training model to label corpus semi-autonomously.Then,we extract various semantic features and build proper feature template.In the end,we utilize CRF to train and test the corpus and extract the sentiment elements.The experimental results show that the method of semi-supervised corpus annotation improved the annotation efficiency.Through combining different feature information,it proved our method outperform the traditional baseline method.Furthermore,after expanding,the sentiment lexicon can extract more sentiment feature and improve the comprehensive accuracy of sentiment elements.Secondly,existing fine-grained sentiment analysis ignores the irrelevant sentiment elements of multi-target and multi-opinion sentences.In order to solve the problem,a method of syntax tree pruning is proposed to delete the irrelevant sentiment elements and extract appraisal expression.To recognize irrelevant comment targets of corpus,we build domain ontology.To delete the syntactic path structure between irrelevant comment targets and opinion words,we build syntactic paths library.With the help of domain ontology and syntactic paths library,we extract comment targets and relevant opinion words,which formed the appraisal expression.Experimental results on three corpuses show the accuracy of sentiment elements can be improved after deleting the irrelevant comment targets and opinion words,and the comprehensive accuracy of appraisal expression also improved about 17%,it proves the syntax tree pruning is effective.