Twitter Voting Predictors Give Mixed Signals
TweetCast is saying Romney. But wait, the needle on the Prez-o-meter is pointing heavily toward Obama. Apparently my tweets are having a political identity crisis.
The basic concept behind these two new sites is straightforward. They are programmed with algorithms that quickly analyze the words used in your Tweets to make a probability-based prediction on how you are going to vote in the upcoming presidential election.
The Prez-o-meter compares tweets to “around 150,000 words each from speeches by U.S. President Barack Obama and Republican presidential candidate Mitt Romney,” according to a blog post about the site. TweetCast is different in that it compares tweets to those of 5,000 voters who have already declared which presidential candidate they will select on Nov. 6.
“It’s not necessarily a useful recommendation or prediction; it’s just a demonstration of this technology,” Shawn O’Banion, co-creator of TweetCast, tells Mashable. “From a research perspective it’s interesting to see the different language that’s being used by the two sides of the spectrum.”
O’Banion is a computer science PhD student at Northwestern University. TweetCast is a project of the school’s Knight Lab, where computer science and journalism students work to create news applications.
A Political Identity Crisis
My personal experience with the websites yielded uncertain results. The Prez-o-meter said I favor Obama by 70%; but when I input the president’s own Twitter handle, @BarackObama, into the meter, it said he only favors himself by 65% (apparently I support Obama more than he does himself). But wait just a second. TweetCast contradicted Prez-o-meter’s diagnosis, and predicted that I would vote for Romney.
While I do tweet about politics, I don’t explicitly favor either campaign. One possible explanation for the confusion is that the language-processing algorithms are not able to account for one of Twitter’s hallmarks: sarcasm.
“Sarcasm is very tricky language,” O’Banion says. “There have been all sorts of efforts to try to detect sarcasm, and it’s really about impossible.”
To test how the sites deal with sarcasm, I entered Stephen Colbert’s handle, @StephenAtHome, into each. Again, there were mixed results. Prez-o-meter had Colbert at 67% for Obama, while TweetCast had Colbert punching his ballot for Romney.
2NIGHT: The Romney-Ryan ticket is firing on all cylinders…a huge improvement for Mitt, who previously was just firing people. TCR,11:30pm
— Stephen Colbert (@StephenAtHome) October 17, 2012
Obama’s war on wealth has me worried. What if my stock portfolio’s hit by a drone strike?
— Stephen Colbert (@StephenAtHome) October 9, 2012
After TweetCast makes its prediction, the tool asks users whether or not the prediction was correct. Since it has only been up for a short while, O’Banion says he hasn’t started fully analyzing the feedback data. His early results seem to say that TweetCast is generally accurate. He has found it to be roughly 80% accurate for people who commonly tweet about politics, and around 65% for those who do not.
These predicting tools are still in their infant stage: Prez-o-meter has been up for about a month, and TweetCast launched a few days ago. Like other probability-based, speech-recognition algorithms, the strength of the tool will improve as it accumulates more feedback data.
What Are the Practical Applications?
O’Banion says he imagines a refined version of TweetCast could be used as a polling tool, or a way to identify moderate voters. Information is a commodity, especially in politics. Just like crops or natural resources, information can be gathered, cleaned, refined and eventually sold for profit.
The Internet and social media sites have led to an incredible proliferation of raw information about people. Going forward, there will be high value placed on any tool that can glean insight from that data.
Karl Klarner, an assistant professor of political science at Indiana State University, is a professional election forecaster. His forecasts are based on historical election trends. He says at this stage, he would not factor these social media tools into his predictions.
“I would be very skeptical of any such thing for the following reason — because it doesn’t have a track record,” Klarner tells Mashable. “It’s kind of a cool curiosity thing, but it doesn’t really have much practical use.”
Klarner shares O’Banion’s opinion that it would be interesting to uncover language patterns behind voting. After making its prediction, TweetCast gives examples of the words, hashtags, mentions and linked websites that helped it make the judgment. For example, Romney supporters are more likely to use words such as “liberal,” “socialism” and “terrorist,” while Obama supporters are more likely to use profanity and words like “womens” and “republicans.”
For now, these sites are simply fun, thought-provoking corners of the Internet. You may enjoy seeing what the sites think about your tweets, and you may even enjoy trying to discover quirks in the sites’ algorithms. With every passing election, however, their language-processing abilities will grow more powerful.
Was the Prez-o-meter and TweetCast able to figure out your tweets, or did the subtext of your 140-character witticisms create a schism in the algoritms? Tell us about your experience in the comments section.
Mashable explores the trends changing politics in 2012 and beyond in Politics Transformed: The High Tech Battle for Your Vote, an in-depth look at how digital media is reshaping democracy.
Read a few of the top posts from the series:
Take it with you, buy Politics Transformed: The High Tech Battle for Your Vote on e-book and get access to four exclusive interviews!