Research And Application On Mining Fine-grained Opinion From Web Comments

As e-commerce platform becomes one part of people's daily life,merchants increasingly hope to understand user's needs and concerns in order to improving their service quality and users are also eager to know the characteristics of products and their performance.Many user comments existing on web provide new way for solving this problem.In this context,this paper conducts a study of mining fine-grained opinion from web comments by applying the theories and techniques,such as natural language processing,data mining,computer programming,etc.,in order to acquire the valuable information such as users' concerns,the feature attributes of products and their performance.This information may provide references on decision-making for merchants and users.The main research of this thesis focuses on the content extraction and opinion sentiments mining from web comments.Different from general opinion content mining,this thesis proposes the synchronous extraction method for extracting the opinion attitude words and commodity feature words in each comment.Based on this,the set of word pair,which is composed of commodity feature words and opinion attitude words,is constructed.By using the word pair set,the valid comments are selected and user's option sentiments in the comments are analyzed.Some achievements are obtained.Finally,by applying the achievements into the real project,the fine-grained opinion mining system based on the network comments is developed under the e-commerce environment.The main research work and innovations of this thesis are summarized as follows.Firstly,by considering the characteristics of web comments texts and the specific requirements of the mining task,the detailed process steps of fine-grained opinion mining work for web comments is given.The main theoretical knowledge involved in this thesis,such as natural language processing,text feature selection,opinion sentiment mining,is researched.Secondly,the method of extracting opinion content from comment texts is researched.The synchronous extraction method is proposed by extracting the fine-grained opinion attitude words and commodity feature words in each comment.A semi-autonomous method for constructing domain sentiment lexicon is designed,which improves the accuracy of the opinion attitude words extraction.The LDA topic model with window constraints is proposed,which uses the position information of opinion attitude words to improve the accuracy of product feature word extraction,and also ensures the synchronous pairing for commodity feature words and opinion attitude words.Thirdly,the problem of mining fine-grained opinions from web comments is analyzed and modeled,which is transformed as one kind of multi-input single-output classification decision problems.The vectorization method of comments text is given.At the same time,the improved algorithm,DFTF,is used to perform dimensionality reduction.Due to the high-dimension and sparseness of the vectorization data,this paper proposes the SVM algorithm to conduct the opinion sentiment prediction.The results from the experiments show that the SVM algorithm has a higher accuracy of prediction and the satisfactory results are obtained.Finally,the fine-grained opinion mining system for web comments is developed and implemented.The overall framework and operation flow of the system are given,and the data storage design based on the MongoDB is discussed.The process of development and implementation of comment collection program,opinion content extraction and excavation program,and result visual display program are introduced in detail.And their key technologies and the final user interface are given.The successful development and actual operation of the system indicates the correctness and effectiveness of the methods proposed by this thesis.