Research On Online Chinese Review Sentiment Classification

With the development of computer and network technology, Internet hasbecome an indispensable information source in everyday life, which has a strongimpact on consumer behavioral patterns. An increasing number of consumers arereading the reviews on the Internet to find out other consumer word-of-mouth ofproducts and services, thereby they can make a good choice. The manufacturers andsellers can also use online reviews as user feedback to learn about the pros and consof their products or services, thereby they can make targeted improvement and gaincompetitive advantage. However, the volume of online reviews is ever increasing.How can online reviews influence product performance, how to deal with theinformation-bearing data to obtain consumer opinion and how to distinguishreviews from non-reviews are problems that challenge the fields of managementand information sciences. Online Chinese review sentiment classification is a newresearch area. As the scale and users of Internet are increasing in China, there is anurgent need of automatically analysis technologies of online Chinese reviews.Based on the existing theories, methods and techniques in the fields ofmanagement science, marketing, natural language processing, text classificationand linguistics, the paper conducts online Chinese review sentiment classificationresearch from three aspects. They are the impact of online reviews on productperformance, calculation of review sentiment orientation and review andnon-review identification. Main contents of the research are as follows.1. The problems of online Chinese review sentiment classification areproposed. Based on the relevant literature of online review mining, this paperdivides them into behavior-oriented and technology-oriented studies and presentsthe three aspects that need further study. The concepts of word-of-mouth, onlineword-of-mouth, online reviews and online review sentiment classification aredefined, and the research scope of this study is determined.2. The impact of online reviews on product performance is investigated. Withdata from Dianping.com, this study uses restaurant pageview as a proxy ofmerchant’s performance and establishes a relationship model between onlinereviews and merchant’s performance. Experimental results show that consumerreviews can positively influence merchant’s performance, but the third-partyreviews and the presence of editor reviews play a negative role on merchant’s performance. The tactics of online review management and usage that third-partywebsites and merchants should adopt are discussed. Finally, the importance ofautomatic sentiment analysis of online review text is proposed.3. The supervised methods of Chinese review classification are studied. First,the vector-space-model-based sentiment classification methods are investigated,mainly including text representation, feature selection and sentiment classificationmethods. Based on N-gram statistical language model, the character-based languagemodel for Chinese review classification is proposed. With English and Chinesereview data, the Na ve Bayes, Support Vector Machine and character-basedlanguage model are compared on different scales of training samples. The impact ofN value in character language model on the performance of Chinese reviewsentiment classification is investigated. The experimental results indicate thatcharacter language model can achieve well performance.4. The semantic orientation-based methods of Chinese review classification arestudied. The Chinese PMI-IR method, especially the impact of search engines andreference word pairs on classification performance is investigated. This studyproposes an approach that utilizes the snippets returned by a search engine toestimate the sentiment orientation of Chinese phrases and reviews, and the resultsshow that performance of the proposed snippet approach is affected by referenceword pairs, windows and classification threshold. Finally, this study compared thePMI-IR and snippet methods in Chinese review classification, and the results showthat with appropriate reference word pair, snippet approach can exceed PMI-IR inChinese review classification.5. A method of automatic bootstrapping creation of Chinese subjective andobjective datasets from unlabeled large-scale Web data is proposed. Thebootstrapping-based model and arithmetic of automatically collecting Chinesesubjective and objective sentences are investigated, and the creation of Chinesesubjective patterns and the identification of Chinese subjective phrases andsentences are the foci. Finally, the defects of the bootstrapping method and thepossible application of the collected subjective and objective data in future work arediscussed.