Research On Fine-grained Sentiment Analysis With Product Reviews

With the development of Internet information,there comes out lots of product reviews in e-commerce.How to organize and summarize these massive information quickly become an urgent problem.Sentiment analysis is based on this demand,through analyzing,sorting and summarizing information automatically to find users' sentiment tendencies.Fine-grained sentiment analysis as the main content of sentiment analysis,the purpose is to extract opinion targets,sentiment words and the sentiment tendencies and so on.Through the fine-grained sentiment analysis,we can find users' satisfaction with products' details,which can play an important role to improve products' quality and find potential users,and also can provide purchasing basis for users.In the study of fine-grained sentiment analysis with product reviews,the extraction of opinion targets and the expansion of sentiment lexicon are the most important tasks.In this thesis,we focus on the problems such as poor portability,a great deal of manual labeling effort,low accuracy in unsupervised learning approaches for the extraction of opinion targets,low accuracy,low coverage of sentiment words and rely on semantic knowledge base in the construction of sentiment lexicon.Specifically,the main work of this thesis is as follows:1.Based on the difference of opinion target statistics across different corpora,aming at the problems of poor portability and a great deal of manual labeling effort in opinion target extraction,an unsupervised approach to opinion target extraction using domain relevance is proposed.This approach first introduces point mutual information to get noun phrases,and extracts candidate targets according to a pre-defined set of syntactic rules.Then use the difference of opinion target statistics across two corpora to get the value of domain relevance for each candidate target.Finally we find out the opinion targets.Experimental results show that the approach can effectively improve the accuracy of opinion target extraction.2.Aming at the problems of low accuracy,low coverage of sentiment words and rely on semantic knowledge base in the traditional approaches,a method of constructing sentiment lexicon based on label propagation is proposed.This approach first selects a certain number of positive and negative sentiment seed words,then uses Word2 Vec to train word embeddings to find out the words that have high similarity with seed words,and then finds out the words that have conjunctive relations with the seed words through the analysis of universal dependencies.Finally we use the label propagation algorithm to determine the polarities of words and get sentiment lexicon.Experimental results show that the approach can construct a more accurate sentiment lexicon.3.Design and implement a fine-grained sentiment analysis protype.The protype can automatically extract the opinion targets from the product reviews,and user's sentiment polarity is displayed in a graphical interface.