August 8, 2011: A masked youth pulls a burning garbage bin set on fire by rioters in Hackney, east London. (Photo: AP/Lefteris Pitarakis)

The hashtags, pictures and quickly tossed off 140-character missives of Twitter could alert police to breaking crises, like riots, before even 911 can raise the alert, according to a new study.

A complex machine-learning algorithm picks up disruptive events minutes, or even an hour, before emergency calls can reach authorities, according to the work of a team from Cardiff University in the United Kingdom.

The team of computer scientists cross-referenced the massive trove of tweets and emergency response timelines during the 2011 riots in London and other U.K. cities, and also the period of unrest in 2015 in the Middle East.

“In this research we show that online social media are becoming the go-to place to report observations of everyday occurrences—including social disorder and terrestrial criminal activity,” said Pete Burnap, a co-author of the study in the journal ACM Transactions on Internet Technology.

The software used by the researchers streams updates using application programming interface, or API. It cuts out extraneous information, removing very short posts and repetitive posts with the same words repeated.

It then ranks the messages, and builds a summary of the overall information dominating the airwaves at a given moment, according to the paper.

The first test event was the Middle East unrest, from Oct. 1, 2015 through Nov. 30, 2015. The dataset included 40 million tweets from 18 million user accounts, which generated 425,000 unique hashtags. That gave a picture of some isolated crimes, transportation updates, politics and weather.

But the real detection implementation was the London riots, incited in response to the police-involved shooting death of Mark Duggan in Tottenham, North London, on Aug. 4, 2011. The riots started two days after the shooting, and spread to multiple U.K. cities and towns.

The trove of 1.6 million tweets spanning from Aug. 6 through Aug. 12 of that year was purchased from the Twitter reseller Gnip, focusing in on hashtags like “#londonriots,” “#tottenham,” “#Londonsburning” and “#riotcleanup,” among others.

Their results: their algorithm generated summaries of impending chaos minutes before the police were notified. For two of the 17 events, the machine was more than an hour earlier to detect rioting in Enfield, based off the “rumour” hashtags.

“Our experiments suggest that our framework yields better performances than many leading approaches in real-time event detections, and using a real-world ground truth published by the Metropolitan Police Services after the 2011 riots in England, we showed our system to detect events far quicker than they were reported to MPS,” they conclude.

“These promising results do not necessarily enable us to ‘predict a riot’ but can provide actionable insights before they were received during the events,” they add.

Limitations hamstring some of the existing tracking technologies, the Cardiff researchers write. The constrained length of the messages prevent some details which would make it more effective—and there are also the challenges of slang, and abbreviated and misspelled words, they add.

A series of programs and tools to scan Twitter for prediction and sensing models of real-world events abound. In the last decade, they have included tools like Twitcident, Twitinfo, Tweetgeist, Tweet4act and Jasmine—coupled often with text-scanning algorithms like LexRank, TextRank and PageRank.

Currently, police have used some full software packages to track social media use in targeted geographic locations. The programs include Geofeedia, MediaSonar and X1 Social Discovery—some of which were used to monitor Black Lives Matter protests and marches. An analysis by the American Civil Liberties Union of North California found that some 500 law enforcement agencies had been using Geofeedia, they reported last year