1Feature-based Customer Review MiningJingye Wang Heng Ren Department of Computer Science Stanford University {johnnyw, hengren}@stanford.edu Abstract The large amount of the information is a big challenge to the customer’s patience to read all the feedbacks. Topical classification and sentimental classification are proposed to be used in information classification. Several machine learning methods, such as the Naive Bayes, Maximum Entropy classification, and Support Vector Machines are good approaches to solve this problem. However, classic sentimental classification does not find what the reviewer liked or disliked. Our approach generalizes an overall rating and user comments on several features for each product. It calculates an overall rating of the product based on PIM-IR algorithm and generalizes these comments on features using feature-based classification. 1 Introduction There have been more and more customer feedbacks on the business website, e.g., amazon, ebay and so forth. Most of them are long and redundant, or even have nothing to do with the product itself. Scanning all of these reviews would be tedious and fruitless. It would be good if these reviews could be preprocessed automatically and customers are provided with the generalized information. Usually a customer-feedback consists of a rating, which is a number, and a quote, a paragraph of judgments. Most current works (Pang and Lee, 2002) are based on the assumption that the rating is binary – good or bad, which is easy to process because of its polarity, but it is not always the case. For example, amazon provides five-star rating, which means that customers could have five choices for rating instead of two. Besides the fact that we loose the polarity here, the meaning of the rating is quite vague. For example, does a three-star rating mean that the product’s quality is not so good in customer’s mind or does it mean that it is not so bad? Moreover, at most time, customers need an overall opinion about the product, rather than a number of particle reviews. That means later customers want to see an overall rating of the product instead of something like ten four-star ratings and ten three-star ratings. There are even more problems with processing the quotes. In the syntactic level, many sentences in review frequently use oral English words or sentence snippets and thus could not be successfully parsed. In the semantic level, many sentences in the feedback are not related to the product itself. For example, they may describe a short interesting story about how they get to know the product, or, they may compare the current product with the others and therefore a lot of information should be related to the other product instead of the current one. There might be some problem with the classic sentimental classification too. It generalizes the opinion of the customer to polarized ones – Excellent or Poor, but it does not find what the reviewer liked or disliked. After all, a2negative sentiment on an object does not imply that the user did not like anything about the product and a positive sentiment does not imply that the user liked everything about the product. Our approach generalizes an overall rating and user comments on several features for each product. It calculates an overall rating of the product based on PIM-IR algorithm and generalizes these comments on features using feature-based classification. Feature Extraction turns out to be a very hard problem and that accounts for the fact that classifying and extracting quotes now are almost done by the human beings – website editors and customers. However, it turns out to be a tedious job. Thus the modern approach is combining the automation with labor force, that is, use automation to extract some potential features and let the human beings check out whether they are really useful. The better result can be get by automation, the less work would be left for the human beings. Here we try to use the feature based sentimental classification to automatically extract features and customers’ opinions toward these features out of the quotes and use PIM algorithm to try to give customers an overall rating of the product. 2 Related Works Most previous research on sentimental classification has been at least partially knowledge-based. Some of this work focuses on classifying the semantic orientation of individual words or phrases, using linguistic heuristics or a pre-selected set of seed words (Hatzivassiloglou and McKeown, 1997; Turney and Littman, 2002). Since these methods take into account the human being subjective factors, some argue that humans may not always have the best intuition for choosing discriminating words than some statistic methods do (Pang and Lee, 2002). However, in our project, it turns out that human subjective opinions would provide more precise result than statistical method. One possible explanation to the contradiction is that the number of samples used in Pang’s result is too small. Pang, et. al also tried to examine whether it suffices to treat sentiment classification simply as a special case of topic-based categorization (with the two “topics” being positive sentiment and negative sentiment). Three standard algorithms: Naive Bayes classification, maximum entropy classification, and support vector machines are tested. It is reported that the results produced via machine learning techniques are quite good in comparison to the human-generated baselines. In terms of relative performance, Naive Bayes tends to do the worst and SVMs tend to do the best, although the differences aren't very large. However, their approach is based on the assumption that users’ opinions are polarized. 3 System Overview Our program consists of seven parts, as the following figure shows. First, we extract the customer feedbacks from the websites and add them to the review database; second, we use POS Tagging to tag these reviews; third, by employing the subject/object review separation described in (Yeh, 2006), we eliminate the objective descriptions, which are not related to the opinion of the customers, from the reviews. Then, on one hand, PMI-IR (Turney, 2002) algorithm is used to calculate the mutual information between the review and the polarized words to generate the weight of the3ratings. The overall rating is the weighted average of each single rating coming along with the product review. On the other hand, we use the