Yes, all very good points. I didn't use a neutral category, as it's difficult to gauge whether a tweet should be neutral. I've thought of detecting them based u

Yes, all very good points. I didn't use a neutral category, as it's difficult to gauge whether a tweet should be neutral. I've thought of detecting them based upon news-speakish tweets, as you've pointed out. That would require its own machine learning run, just to sort out neutral from containing-sentiment. Also keep in mind, different topics work better. The term "election" has a lot of news headlines, which many are probably neutral, skewing the results. More consumer-ish topics yield better results. But yes, tweets are difficult to analyze. I've done another recent experiment with tweet analysis, if you like this kind of stuff http://primaryobjects.com/CMS/Article158.aspx http://primaryobjects.com/CMS/Article158.aspx