Hey there, drunk tweeters. We see what you did there — and so do scientists.
A team of researchers led by Nabil Hossain at the University of Rochester have developed software that can scan Twitter for boozed-up messages.
To develop the algorithm, researchers first collected a year's worth of geo-tagged tweets — numbering in the millions — from Monroe County (in upstate New York) and New York City, and then filtered them for references to drinking and alcohol.
The information they gleaned from those tweets was then used to teach machines to recognize which ones were about being drunk or drinking, and those sent from people who were likely and actually drunk at the time.
That initial sort job was no match for computers, though, so Hossain and his colleagues sent them to Amazon Mechanical Turk — a site that crowd-sources human labor. A worker would look at a tweet to decide if it referenced alcohol. If it did, they determined whether the tweet was about that Twitter user or another person, and if the tweet happened in the moment of intoxication.
Tweets that passed all three tests fed an algorithm that learned to distinguish drunk tweets from totally sober tweets about drinking.
There were some unexpected, if not inexplicable, findings. For example, tweets with "dad" in the text seemed more likely about to be about the user drinking; meanwhile, tweets with the phrase "my mom" seemed more likely to be about other people drinking.
Here's one table showing which words may be more highly correlated with tweeting-while-drinking:
The researchers also wanted to know where people drank relative to their homes, so they performed a similar process on a set of geo-located tweets with phrases associated with being home ("home," "bath," etc.). Using both methods, the authors claim they can pinpoint a person's home within 300 feet up to 80% accuracy.
So what does this all teach us?
For one thing, Twitter users in New York City had a much higher rate of tweeting-while-drinking near home compared to Monroe County residents, who did so farther away from home. As the authors point out, this might have a lot to with how dense NYC is, not only with people but also access to booze. (In Monroe County, it's generally a farther hike to a bar.)
The researchers also tracked areas with a higher proportion of tweets sent while drinking compared to other geo-tagged tweets, which they call "unusual drinking zones."
Here are where Twitter users are drunk-tweeting at higher rates than the rest of NYC:
Hossain emphasized that this isn't just fun and games: The Centers for Disease Control estimates that alcohol use causes roughly 88,000 deaths each year, resulting in 1 in 10 deaths among adults aged 20 - 64.
"I'm mainly excited about approach that can help in reducing the damages of alcohol use in the community," he wrote in an email to Tech Insider.
Hossain added that future applications could come in the form of a mobile app that could alert a "helper" when a friend is intoxicated nearby.
He also said he's curious how the social media landscape and drinking culture might intersect and inform each another.
To demonstrate what the software can do, Hossain shared a few tweets with us that the algorithm identified as "tweeting while drinking." The results certainly seem impressive:
The study fully recognizes that Twitter users are a very small subset of the population and not representative of the general population. But the level of detail that social researchers can gain from social media is still mind-boggling.
Hossain also told us the algorithm accurately located a neighborhood popular for drinking among students at the University of Rochester all by itself.
That's artificial intelligence that knows where the party's at.
