Think last night's drunk tweets were pretty coherent? You won't fool University of Rochester researchers, who have developed a machine-learning algorithm that can tell when a tweeter is drinking. To do so, they started with humans: Researchers collected tweets associated with alcohol—think ones with words like "drunk" and "party"—sent from NYC or New York's Monroe County between July 2013 and July 2014, reports the MIT Technology Review. Workers with Amazon's Mechanical Turk service then helped determine if a tweet mentioned alcohol; if the tweeter himself/herself was the one drinking; and, if yes, if the tweet was posted while drinking. Next, researchers compiled a separate selection of geotagged tweets that featured one of 50 home-related keywords, like "sofa" and "bath." Mechanical Turk workers were then asked to determine if each of those tweets was sent from home.
The results provided enough data with which to create a machine-learning algorithm that's able to reveal "patterns of alcohol-related behavior in unprecedented detail," reports the Technology Review. Researchers noticed a positive link between the density of liquor stores and bars and the amount of tweets about drinking. They also found NYCers sent more tweets associated with alcohol, and more often did so from home. Those in Monroe County were more likely to travel at least a half-mile from home to drink. Indeed, "tweets can provide powerful and fine-grained cues of activities going on in cities," researchers say, per the Christian Science Monitor. And while the youthful nature of Twitter could skew data, the Technology Review observes there's "great power" in this approach, which is far less costly and cumbersome than, say, surveying people about drinking via a questionnaire. (Scientists say there are four kinds of drunks.)