Observer: Opinion Analyzing and Comparing Opinions on the Web
Bing Liu, Minging Hu, Junsheng Cheng
Presentation by Mark Strohmaier
Problem Overview
The Internet allows people to see product reviews from a large number of people
Often easier for customers to compare based on features rather than reading many reviews
Easier for companies to improve products if they know what aspects people liked and did not like
Problem Overview
Challenges
Locate features which users rated Identify whether the rating was positive or negative
Further complicated by three review formats Pros and Cons (Cnet) Pros, cons and review (Epinions) Free-form (Amazon)
Problem Description
Each product Pi has a collection of reviews Ri Each review contains Product Features attribute of the product mentions review Opinion set whether review treats the feature positively or negatively
Positive (negative) opinion set set of all positive (negative) reviews of a product feature
Problem Description
Columns show: 1) Actual number of scores 2) Percentage of Scores
Key assumption
Extracting Features
Determine part of speech (using linguistic parser) Replace words with [feature] Use 3-grams to make shorter segments Distinguish duplicate tags, reduce words to stems Generate rules using CBA association mining system
Processing Rules
Only keep rules which end in [feature] Remove rules not sufficiently predictive Generate language patterns
System Architecture
Preliminary Results
Conclusions
Unique visual approach
Supervised pattern discovery method