I'd just use Weka to train a classifier based on a batch of manually-sorted entries. Check Naive Bayes and K-means to start. The result will be more specific to

I'd just use Weka to train a classifier based on a batch of manually-sorted entries. Check Naive Bayes and K-means to start. The result will be more specific to your content and you'll be more able to refine it later to improve and optimize the error rates. Sentiment is fiddly, hard to grasp. Trying for fine gradations is also the road to lots of errors. I would stick to positive, negative, neither. While you're at it, there has been some work done with distinguishing valid vs. fake reviews. http://portal.acm.org/citation.cfm?id=1612762.1612776 http://portal.acm.org/citation.cfm?id=1612762.1612776