| Review is a kind of subjective text which reflecting the value of things. It has beenwidely used in product attribute extraction, preference learning and opinion analysis andmining. However, reviews can be numerous and varying in quality. So, it is very importantto automatically estimate the quality of product reviews (such as "expert" level review or"junk" review), which are of great value to review-based researches and applications. Thereview quality assessing has been regarded as a task of classification or ranking.Quality-based classification aims at distinguishing the "high quality" reviews from allreviews, and review ranking sorts reviews of a certain product by their quality. Bothclassification and ranking are supervised learning methods. Thus, feature extraction andreview quality modeling are two major studies in this thesis.From the perspective of the user’s intention, we explored factors which affect thequality of a review. The review quality is directly judged by users. So it is quite natural toexplore the features which reflect the user’s intent. Besides, these features are usually ofgreater versatility. From the point of the users, we proposed three factors which reflect thequality of reviews, which are Information Needs Fulfillments, Sentiment and Opinions ofReviews and Information Credibility. Each of these factors was measured by some metricsand the corresponding mining algorithms were proposed. Experimental results show thatthe user’s intention based features have a good effect on review quality assessing task.After combining our features and the features used in previous work, we obtained betterperformance.Besides the new features, three review quality models are also built from different perspectives, which namely: Time Quality Model, Information Overload Quality Modeland Topic Quality Model. These three models are independent of each other. Finally, thesethree models are used for optimizing the quality of ranking system, and achieved highperformance. |
PDF Full Text Request