I was just having been assigned a project of conducting sentiment analysis for some document collections. By Googling, a lot of sentiment-related research has popped up.
My questions are:
What are the major methods/algorithms for sentiment analysis in the field of machine learning and statistical analysis?
Are there any well-established results?
Are there any existing open-source software that can perform the sentiment analysis?
Try SentiStrength which performs well compared to similar algorithms, and the associated research papers. Discussion of other tools and methods can be found here and here.
I have the impression that much of what is being done here is extremely heuristic. In fact, most people seem to apply this to the <120 chars of twitter statements. Probably the results (while not being computed this way) aren't much better than counting "positive" and "negative" words with a litte position information ("A better than B" = positive for A, negative for B)
When you then see companies buying a full twitter feed (that's how many mbit per second?) and claiming to do sentiment analysis on this this seriously makes me wonder if there is any statistical validity here. No wonder e.g. Yahoo failed badly at predicting the preelections for South Carolina: http://www.technologyreview.com/web/39487/
People are way to proud and keen on just being at all able to process the amount of data, they completely seem to neglect properly validating their performance.
Sorry to be this pessimistic about the state of the art.
