On a recent rainy Thursday night, a couple hundred people gather after hours at the ECHO Lake Aquarium and Science Center to listen to a University of Vermont assistant professor of math and statistics talk about happiness — or, more specifically, using Twitter and other online communications as a barometer of happiness.
Chris Danforth is an engaging public speaker who doesn’t need any charts or graphs to articulate the findings of his big data project. The crowd at ECHO will leave tonight knowing this: There’s a lot to learn from social media and digital communication. You just need to know how to harvest the data.
Danforth works at UVM’s Vermont Advanced Computing Center, a research facility that’s essentially about using computers to do more than browse eBay and send email. Along with fellow professor Peter Dodds and a team of undergrad and graduate students, Danforth is using those high-powered computers to glean data from billions of tweets sent from all over the world. Computer algorithms and Twitter’s open-programming interface allow the researchers to amass huge amounts of raw data. They then use their human imaginations to figure out which information to look at.
The team’s research into Twitter addresses the question: How can we measure happiness? Danforth notes that gross domestic product is not necessarily the best way to measure the collective well-being of our society; Bhutan has actually adopted the concept of “gross national happiness” to gauge the total health of the country. But happiness is a lot harder to quantify than are most economic indicators. So Danforth and his team have come up with the “hedonometer” to measure happiness through the words we use. Their research isn’t limited to Twitter; they pull content from blog posts, music lyrics and New York Times stories, to name a few other sources.
“What we’re doing is collecting words. We collect every public Twitter message that Twitter will give us,” Danforth says. The messages — about 20 million English-language tweets a day — come from an open application programming interface that Twitter makes available to developers. “What we do is watch trends in the way different words are used,” Danforth continues. “Underlying all this is the assumption that each word has some happiness value associated with it.”
How do the researchers determine the happiness value of an individual word? The team turned to a website called Mechanical Turk, a division of Amazon.com, which facilitates paying large groups of people a small amount to complete menial tasks. In this case, the researchers had 50 people assign happiness scores to the 10,000 most frequently used words on Twitter. The scale is simple, ranging from 1 for the saddest words to 9 for the happiest. The average of those 50 scores becomes that word’s happiness value for the purposes of the research.
This is where the data start to become more revealing. The “happiness values” of those 10,000 most frequently used words are a good starting point, but you don’t really need to survey people to assign a word’s happiness value. “The collection of sentiments expressed about a particular topic tend to be very closely related to the sentiment of that topic itself,” Danforth says.
Say the topic is “cold,” and the researchers want to assess the happiness value of the “bag of words” in tweets that contain that word. “Shake out all the ‘colds’, because I know that’s in every one of them,” Danforth says. “And then I compute the happiness of [the remaining words]. Turns out we get a number that’s very close to the number that the Amazon people gave to the word ‘cold.’”
This finding means the researchers can gauge the sentiment of words that weren’t among the original 10,000 but become more common over time. Like, say, “occupy.” “I don’t need to score ‘occupy,’ because I have scores for the billions of words [people use with ‘occupy’] … and the average of them all will be very close to the score that ‘occupy’ gets,” Danforth says. “And that will change over time, because two or three months ago it didn’t mean anything.” This means the happiness scores are not absolute, but act as a reflection of real-time sentiment about a given word. A word like “happiness” won’t see its score change too often, but words tied to news and current events could vary wildly.
This is one of the most illuminating parts of the project. UVM graduate student Kameron Harris illustrates how words and phrases can see dramatic shifts in sentiment over time. Golfer Tiger Woods used to be one of the most revered figures on the internet, but his score tanked in November 2009 when word of his marital indiscretions leaked out. By contrast, public sentiment toward Michael Jackson improved after the singer died in 2009, as the prevailing associations with him shifted from allegations of child abuse and strange behavior to memories of his great music.
Harris says he’s combining linguistic and geographic data to see how tweets containing the word “Irene” have shifted in Vermont — perhaps becoming more positive as tweets about the storm’s damage gave way to ones about communities pulling together and rebuilding.
The researchers are also looking at every blog post that uses the words “I feel” or “I am feeling,” using data acquired by digital artists Jonathan Harris (a Shelburne native) and Sep Kamvar for their online data visualization project “We Feel Fine.” In that information, Danforth and his colleagues found another surprise.
Many bloggers post their age, making it possible for researchers to correlate that with words and happiness. Mashing up age data and happiness scores, they found a mountain-shaped bell curve revealing that younger and older people are least happy, while people in midlife are the happiest. This is the opposite of results researchers have attained in old-fashioned phone surveys, where a random sample of people were simply asked how they felt.
Danforth says this project could provide a “more honest assessment” than a survey of participants, “because it’s [people’s] behavior as opposed to their reported behavior,” he notes.
“Whenever you ask somebody a question, all kinds of complicated stuff happens in their brain,” Danforth says. “But if you’re just looking at what they’re doing, and hopefully doing it in a way where you’re not invading their privacy ... then hopefully you’ll be able to learn something more about them than if you asked them the question.”
(Don’t worry, bloggers and Twitter users, Danforth doesn’t have a happiness file on you; the data his researchers use are all anonymized.)
Ultimately, the UVM research goes beyond a real-time measurement of happiness and becomes a larger study of how we use language. Danforth says he and his team expected the distribution of happiness values for all words to be roughly normal, a bell curve centered over the middle of the 1 through 9 happiness scale. While they did find a bell curve, its center turns out to skew a full point toward the happier side of the scale, centered over 6. This is true of all their sources of data — tweets, blog posts, even New York Times stories of the past 20 years.
“It appears there’s a bias toward happiness that’s kind of built into the way we communicate with each other,” Danforth says. “It’s not an accident that our language evolved to be like this.” He and Harris haven’t yet determined if a similar bias toward happiness exists in other languages.
Harris points out that there could be psychology at work here, too. “When you make a negative statement, you try to make it more positive,” he says. Essentially, our beating around the bush skews the results toward the happy end of the scale.
In the future, the researchers plan to continue the project with even more data sources, such as Google Books, the internet giant’s searchable digital database of literature. Harris says they would also like to create a website that would give the public an interactive way to explore their data.
Sure, some of the findings may seem obvious. It’s probably no surprise to anyone that people thought less of Tiger Woods when he was revealed as a serial cheater. But happiness and other emotions have always been nearly impossible to quantify, and, as Danforth puts it, “It’s the things that are hard to measure that are really important.”
Guess we can learn a lot more from Twitter than what our friends are eating for lunch.
Watch Danforth's talk on Vermont Public Television's website.
See the results of the team's research at onehappybird.com.