Re:Show HN: Tracking voter discussion of the Presidential de...
by erikrose on Wednesday October 03, 2012 @10:10PM
(#4610590)
What you're talking about in the last paragraph is sentiment analysis. This is really cool stuff, but it's tricky. Even if you have 2 humans do sentiment analysis independently, they agree with each other only 70% of the time. Computers tend to do worse, of course.
What makes it extra challenging is that tweets are short, so you have wide error bars on the sorts of math you'd depend on in larger corpuses. I think you'd have to actually bang a parser against it and try to understand what the tweeter is saying.
Re:Show HN: Tracking voter discussion of the Presidential de...
by erikrose on Wednesday October 03, 2012 @10:12PM
(#4610592)
Btw, here's an example of somebody trying to do sentiment analysis on tweets: http://www.sentiment140.com/search?hl=en&query=romney.
Fascinating stuff, I also noted a few posts positive to romney that wer associated as obama as obama was mentioned and vice a versa. I suspect a first mentioned corralation, though have seen exceptions to that, i'd paste examples but no pause button and late. To truely seperate those type of posts is something were you will need to crowdsource down the line if only to learn a rule set. Though tweets are short so permutations will not be that dynamic as apposed to more open social means and with that the limitation can only help.
Anybody aware of a ruleset that can weight one person over another based upon the content of block of text that mentions both people. Certainly would have uses beyond this. Amazing how something so simple as that involves alot more than you would initialy think.