The study focused on more than 530,000 tweets collected between March 14 and March 24. Ramesh and Seetharam tracked popular hashtags and topics, and collated them into six general categories, with some tweets landing in two or more. The categories included:
General coronavirus tweets (with hashtags such as #COVID19, #outbreak, #coronaapocalypse and #pandemic).
Quarantine and social distancing (including #QuarantineLife, #SocialDistancingNow, #flattenthecurve and #workfromhome).
School closures (#closenyschools, #suny, #cuny and #homeschool2020).
Panic-buying (#PanicShopping, #toiletpaper, #handsanitizer, #coronashopping and #WashYourHands).
Lockdowns of cities and states (#lockdown, #Shutdown and various region-specific variations).
Frustration or hope related to the pandemic (#CancelRent, #fightcorona, #saveyourlife, #COVIDIOTS, #DrFauci, #stopthespread, #saveworkers, #WhenThisIsAllOver and #StaySafe).
Even looking at just a 10-day period, the Watson researchers found the crisis “evolved quickly” from one topic to the next.
“Around the middle of March, we saw a boom in panic-buying of essential goods,” said Seetharam, who leads Watson’s future internet design lab. “The first thing that ran out was toilet paper, so there was a lot of talk on Twitter about that. Then those topics went away because schools closed and people figured out that the crisis about essential goods would persist for some time … After schools closed, we saw how the infection was spreading through tweets about social distancing and shutting down bars, cities, states and even hashtags like #ShutDownUSA asking to shut down the whole country.”
Watson refers to the Thomas J. Watson School of Engineering and Applied Science.
Across all the categories, some common words appeared frequently, among them death, virus, cases, family and health. Ramesh and Seetharam stripped out other words, like place names, because they were not relevant to the study.
Although the researchers looked at larger trends, they also were able to zero in on more personal experiences.
“One tweet said, ‘Teachers are parents, too,’” said Ramesh, who runs Watson’s machine learning research group. “It exemplifies how people are showing emotion during this time — how they are feeling for the other person.”
Getting the Twitter data requires a request to the social-media company, and researchers can receive only about 10 percent of tweets from a particular time frame, but that amount is considered a “valid representative sample.”
Ramesh and Seetharam are continuing to collect info for a more comprehensive study in the future.