Computer program uses Twitter to 'map mood of nation'
bbc.co.uk
bbc.co.uk
For instance, you have to filter out most posts with links (because they tend to have very little associated data, or are spam), anything in languages you don't have parsers or datasets for, etc. Then you can cut on character count (because there are a surprising amount of useless tweets that are under 5-9 characters)...
As others have and will continue to state, this further cuts down the sample size of an already fairly small sample (500+ million in the world[0]). If you want to be specific and only look at a single nation, you're cutting that sample size based on who has location turned on and/or broadcasts it in their profile. This is a significantly smaller number.
Edit: This is obviously a better sample size than a random poll of individuals, however, but that doesn't mean that the cumulative quality will be drastically higher.
[0] - http://www.statisticbrain.com/twitter-statistics/ (sources from three different input points)
Although, I'm not classifying by specific emotion type as the article describes. In all, I'm seeing accuracy rates of around 80%. This is in comparison to a brute-force word list, which scored an accuracy of 62%. I'd be curious to see what kind of accuracy the researches in the article are achieving, with so many sub-division classifications. Not to mention, how did they derive their initial model?
After the presidential election, Gallup overhauled its polling methodology. Why? Because they significantly overestimated Mitt Romney's popularity, and one of the factors may have been relying too much on landline polling, as older (i.e. more conservative) people are disproportionately likely to still use landlines.
If Gallup's predictions are as sensitive to a factor like that, I imagine this neophyte tech/science company is going to be way off in their measurements.