I read some papers and post these days about product review mining. I want share some ideas of mining I summarized.
First, This mining is a kind of text mining.The review mining also call opinion mining. The researchers want find what are the reviewer's opinions on products, either negative or positive.
The review mining do not focus on the ratings, such as product rating on amazon. This mining focuses on the text written by customers or professional reviewer. Certainly, the mining result is helpful for producer to improve products. If the reviews are classified by "cons" or "pros" automatically, such as Newegg.com's review, it is much easier to mining.
The mining is to find feature words, and then based on the number of feature words, measurement scores are calculated.
Feature words could represent customer opinion directly. For example, if customers say "awesome", "excellent", these words show their positive opinions. But if they say, "bad", "s***", negative opinions are shown. Also some feature words are depend on different products, for instance, most customer need a quiet computer case, so "quiet", "no sound" are good for computer cases, but for stereos, these words are negative.
To find the feature word is regular text mining. After searching the feature word, some evaluation method are developed to decide whether the comment is negative or positive The snip below is used ref-3, the SO value is standard whether this positive or negative.
Ref:
1. http://www.slideshare.net/felipemattosinho/mining-product-opinions-and-reviews-on-the-web
2. Movie Review Mining and Summarization, DOI: 10.1145/1183614.1183625
3. Movie Review Mining: a Comparison between Supervised and Unsupervised Classification Approaches, DOI: 10.1109/HICSS.2005.445
While feature words are key for review mining, I feel that incorporating the actual ratings would prove prove beneficial. I'm not saying that these companies aren't already doing it, but just in response to what you posted. Say you have this algorithm to find feature words. What if this site had particularly lengthier reviews. For example: A negative review without any negative feature words, but it is negative due to a comparison being made to a better product, which good feature words are being used for. Of course, this is assuming that the company doesn't already have an intense machine learning algorithm in place. This is where I see combining ratings with review mining would yield a more complete picture of one's review.
ReplyDelete