This is not how sentiment analysis work (or should work). I worked on something similar (Naive Bayes based sentiment analyzer https://github.com/mohitranka/Twi

This is not how sentiment analysis work (or should work). I worked on something similar (Naive Bayes based sentiment analyzer https://github.com/mohitranka/TwitterSentiment https://github.com/mohitranka/TwitterSentiment). I also work for a company which is in the same space as groubalcsi.com (Brand/Product opinion mining) Sentiment analysis is not a classification problem (like spam detection), but it is an identification problem, because sentiments are always associated with an entity (and attribute, if specified). For example, a tweet saying, "Dell is not as good as apple" requires to identify entities (Dell and Apple) and associate sentiments to them (Negative and Positive, respectively). It is incorrect to try to associate sentiment (whatever it may be) to the tweet itself.