[mweatherill]

The fundamental problem with social media sentiment analysis is that it is based upon the assumption that tweets are representative of the population. It ignores the issue that a large percentage of tweets come from spam bots.

As soon as you start measuring sentiment, there is an incentive for an interested party to try sway the results. For example, I could use a spam network to publish opinionated tweets just to obtain a headline in the media (E.g., "95% of tweets agreed with X"). The media coverage gives legitimacy to the manipulated result.

I wonder how many companies have set goals around this sort of sentiment. "Let's increase positive sentiment by 10%". An ethically challenged consultant could easily manipulate results to meet the stated goal.

There is still a place for sentiment analysis and that is in targeting individuals for follow-up. Assuming that the user can be matched to a real consumer, this avoids the issues of statistical manipulation. Though users can still exploit the "complain and get something free" loophole.

[burke]

Nevermind that even with spam bots filtered out, twitter users are not nearly representative of the voting base.

[gklitt]

This is an important point to keep in mind. I was surprised that the analysis seems to suggest a slightly more favorable view of Obama's performance than Romney's on Twitter, given the general consensus in the media that Romney won. It's probably because the young, plugged in Twitter crowd leans to the left.

[lobotryas]

Good point. I'd say the answer is to try and obtain access to more data. Imagine if you had access to the Facebook "firehose" and could analyse every single FB message, wall post, Like and so on. I would say that Facebook is a better representative of the general population than Twitter and with time you could tune the model to provide insights into user sentiment. Would be cool to see what the userbase thinks about X brand and if that brand wants to pay in order to boost its user sentiment on FB.