- Luke Eastman
Roughly 500 million times a day, someone in the world writes a tweet. Every second, nearly 6,000 bursts of text, image and emoji, 280 characters or less, are sent out into the world. About one-third of them are in English, but users worldwide use Twitter in many languages to share their feelings, food, pets, work, classes, relationships and nearly every other aspect of their lives. For every funny dog picture, there's an ill-informed yet confidently delivered opinion; for every news article, a joke.
Twelve years ago, researchers at the University of Vermont's Computational Story Lab designed a use for all those billions of tweets: the Hedonometer, software that uses tweets to track global happiness. What the Hedonometer has confirmed about the last two months probably won't surprise you: We've all been really bummed out.
The UVM researchers, led by Peter Dodds and Chris Danforth, struck a deal with Twitter to scrape a random 10 percent of all tweets daily. In 2008, when they started, this was a novel idea, and the tweets came at a rate more akin to a trickle than today's fire hose. The lab's backlog now contains more than 10 billion tweets in English alone and spans more than 180 languages.
The Hedonometer tracks 10 of those languages. Researchers crowdsourced rankings of 10,000 unique words on a happiness scale. The happiest English words are pretty obvious — "laughter," "happiness," "love." People are also pretty happy when they're talking about vacations or family members. Prior to the pandemic, the unhappiest words included "terrorist," "cancer" and "suicide."
But the coronavirus has had an unprecedented effect on the Hedonometer across languages. Usually, when something bad happens, there's a dip in global happiness that lasts about a day before our collective attention moves on. Holidays tend to cause similarly brief but dramatic happiness upticks.
On March 11, 2020, Tom Hanks announced he had the coronavirus, the National Basketball Association canceled its season, President Donald Trump banned European travel to the U.S., and stock markets plummeted. The following day was the saddest in Hedonometer history.
"There've been lots of sad days on Twitter. There's these big dips when a celebrity dies or there's a natural disaster or a mass shooting," Danforth said. "But there's never, in the entire history of our instrument, been a sustained sadness like we've been seeing."
Every day in March, he said, has been collectively sadder than the day of the Boston Marathon bombing in 2013. And the saddest words on Twitter? That list is now dominated by pandemic-related terms such as "ventilator," "sanitizer" and "quarantine."
Computational Story Lab research can tell us about more than our global mood, though. The lab has produced a wide variety of reports based on analyses of tweets, which researchers are now using to better understand how global society is responding to and processing the coronavirus pandemic. Those researchers can tell you, for example, exactly when most of the world was and wasn't paying attention to the coronavirus.
"The world's collective attention dropped away as the virus spread out from China," said David Dewhurst, a UVM research fellow who's also a data scientist in the private sector. Virus-related words peaked in late January and early February, "and then they sort of fall out of favor. People are talking about other stuff for a while."
Interested in those results, Dewhurst thought to take it a step further. He split the pandemic-related words into two clusters and found that more concept-based words associated with pandemics, such as "epidemiology," dominated the online conversation in the earliest days of the pandemic. But as the coronavirus spread, discussion of it became much more specific, with words such as "flatten" and "distancing" more strongly represented.
Dewhurst and the other researchers examined the volatility in usage of the concept-based words. "Volatility," as Dewhurst uses it, refers to the magnitude of the percent change in how frequently a word is used. They found that the volatility in usage on one day could be associated with a similar volatility in the number of coronavirus cases roughly 23 days later. In other words, the number of people tweeting certain words did not correspond directly to the number of cases; rather, the change in the usage of those words corresponded to the later change in the number of coronavirus cases. This association held true across 24 languages that Dewhurst examined.
To be clear, association doesn't mean prediction. Dewhurst emphasized this point because, he said, similar research has started to emerge elsewhere, and scientists and media alike are suggesting that social media can predict incidences of the coronavirus. The online news source Business Insider, for example, wrote a story in early April about research from a firm called Dataminr that claimed to predict 14 states where infections would spike next, according to social media activity.
"Saying that something predicts something else is a very hard thing to do," Dewhurst said. He hopes his research "could have predictive power" at some point in the future, but he doesn't imagine it will do much to help us address the coronavirus pandemic, which is already embedded so deeply in our lives.
"There will be future pandemics. This is something that we know," Dewhurst said. "One of the things that interested us is that the words that were associated with future percent changes weren't specific to coronavirus."
Rarely do we face something as unpredictable as this pandemic, Danforth said. When the impending disaster is a hurricane, for example, we can launch satellites and observe storm patterns and predict, with reasonable accuracy, how soon the hurricane will make landfall. Most importantly, our increased attention on the hurricane doesn't affect its path.
"The hurricane doesn't change its track because of the forecast. And we, in this epidemic, have changed our behavior quite dramatically," Danforth said. "When you make predictions about a social system, as opposed to a physical or technological one, there's a lot more uncertainty."
It's not all bad news out of the lab. In recent days, the Hedonometer's global happiness level has slowly crept back up. We're now at the happiness level of March 7 — which, to this reporter, feels like a lifetime ago.
Maybe we're just adjusting to our new normal, or maybe things are actually looking up. One thing's for sure: Wherever this pandemic goes from here, we'll be tweeting about it.