Everybody hated the How I Met Your Mother finale, right? Well, maybe not. The hour-long conclusion to the popular CBS sitcom was the target of much derision after it aired March 31. (See: Here, here, here, here, and here.)
But data, as it sometimes does, tells a different story. Most of the people surveyed by the social analytics platform Canvs felt good about the finale. More than twice as many people either enjoyed it or called it "great" than those who said it "sucked."
Canvs, which formally launched last week, uses sentiment analysis to figure out how people feel about TV shows and other entertainment. That analysis is then combined with scary-specific data profiling, so Canvs can tell its clients not just how people feel about what they're watching but what kinds of people feel certain ways.
The viewers who thought the How I Met Your Mother finale was "sad," for example, were also more likely than the average Twitter user to be teenaged girls interested in Taco Bell. More on that in a minute.
First, here's what Canvs found, based on Twitter reaction from about 117,000 viewers: More than half of reactions to the show were positive, while about one-fifth of them were clearly negative.
Here's the breakdown:
• 19 percent of people said the finale "sucked, not a fan"
• 18 percent enjoyed it
• 16 percent called it "great, glorious"
• 15 percent called it sad
• 13 percent reacted with "love, perfect"
• 12 percent "loved(d)" it
• 7 percent called it "great, sad"
As with any dataset, there are limitations. In this case, we're talking about 117,000 tweeting viewers—a bigger sample size than many telephone-based surveys, but still only 1 percent of the 12.9 million people who watched the finale. (To put that in context, though, presidential campaign polling routinely is based on sample sizes that represent 0.0001 percent of the actual population.) The Canvs analysis covered about 185,000 tweets, just a portion of the half-a-million tweets that Canvs identified as being related to the finale. The platform only analyzes tweets it is sure it can interpret accurately, founder and CEO Jared Feldman told me.
Feldman says his team obsesses over how we talk about what we feel. Computers aren't always so discerning, which makes much sentiment analysis flawed. An algorithm might recognize the word "enjoy" in a tweet that says, I really didn't enjoy the How I Met Your Mother finale, without realizing that the tweet isn't ultimately positive. You can teach a computer to recognize the "didn't" before "enjoy," but that doesn't go far enough, either.
Understanding how humans talk about their emotions is far more complex than accounting for qualifiers like "don't" and "can't," which is part of what makes machine-run sentiment analysis so thorny.
"Maybe you said, 'I can't effing stand how much I love this show,'" Feldman told me. "The computer's mind is blown... The way that 12-year-olds talk about loving Justin Bieber? There's no dictionary on the planet that captures that."
So Canvs uses an algorithm built on years of more nuanced human analysis known as "supervised sentiment analysis" in the industry. The result, Feldman says, is real-time conclusions at a level of sophistication that previously would have taken hours. Sentiment analysis is notoriously difficult to get right, and it's not really possible to tell from Canvs' interface whether the platform is as nuanced as Feldman says. It's user friendly and incredibly detailed, but using it requires faith in an unseen algorithm—one that Canvs is selling.
But Canvs gives its users plenty of opportunity to assess the data. For each reaction category—like "sucked, not a fan," or "great, glorious,"—you can dig more deeply into what people actually said. Canvs displays popular keywords from related tweets, user's handles, and the tweets themselves. Here's a glimpse of the breakdown of people who reacted with "enjoy:"
The graph lets you track minute-by-minute reactions throughout the show's airing, so you can see when "sad" spiked, for instance.
This analysis becomes valuable for companies that want demographic profiles that include how people feel. So Canvs reviews months of backdata—what you've tweeted, who you follow, where you're tweeting from—to deduce personal information like a viewer's age, gender, ethnicity, income levels, interests, brand loyalty, and so on. We already know that marketers are combing the internet for hints about customers, but Canvs shows how easy it is to get a detailed picture. Here's a glimpse at the people who thought the finale "sucked:"