I think you're on to something about bad categorization. Sentiment analysis as practiced is almost always pseudoscience. (And that's a hedge. I've never seen it done right but I'm not outright discounting the possibility.)