Mining on-line user reviews
Bin Shi · 2007
In this report, we propose a system for mining on-line user product reviews.Each sentence in a review is analyzed to extract the user's opinion/orientation of a product feature.The sentiment/orientation towards each product feature is then tallied across all users to generate an overall scoring for the product and its individual features.For testing purposes, we extract reviews from the Chinese cellphone on-line community.On-line product review mining falls under the realm of web structured data mining.In order to unify the different variations describing a particular feature like screen/display, we manually created an ontology of unique product features specific to cellphone with a list of commonly used synonyms.Later on, this ontology is enhanced by statistic analysis results on on-line cellphone reviews.After surveying the information extraction research landscape, we decided on using a template-based hybrid extraction method.We proposed using a novel method to rid Chinese content from noise such as extraneous punctuation frequently used in on-line posts.Sentiment classification has attracted quite an interest in the community; we therefore first surveyed existing sentiment classification techniques.Next, an improved lexicon based sentiment classification method and an machine learning technique based sentiment classification method, both operating at the sentence segment resolution, are proposed.In order to present intuitive review results to the end user, we have created an interactive graphical web interface which allows the user to show/hide/aggregate selective review feature scores.A fully-functioning prototype has been created to demonstrate the end-to-end system.Experiments were conducted to evaluate our suite of proposed methods.Based on analysis i