Jiho Han Ronny (Dowon) Ko Objective: automatically generate the summary of review extracting the strength/weakness of the product Use NLP techniques to predict ratings ◦ Similar to sentimental analysis Key Insight: Imposing market structure assumption ◦ Different type of information extraction Amazon review text Opinion = (orientation, polarity) Review Texts Orientation Profile ∞ m k Rating Parsing – through Stanford NLP syntax parser Initializing orientation and polarity ◦ Selecting polarity words through decision tree (Max-Ent) ◦ Orientation using N-gram (uni + bi) ◦ Use wordnet when testing Extract market profiling and pricing kernel Update word polarity Repeat until no more improvement Extract the words that have significant effect on rating (in terms of maximizing entropy) Initial word polarity Change in polarity Performance