Nieman Foundation at Harvard
Postcards and laundromat visits: The Texas Tribune audience team experiments with IRL distribution
ABOUT                    SUBSCRIBE
Jan. 23, 2013, 10 a.m.

Feelings, nothing more than feelings: The measured rise of sentiment analysis in journalism

Tools to derive opinions from social media postings are far from perfect, but they also offer hope for getting beyond pundits claiming to speak for others.

“Punditry is fundamentally useless,” Nate Silver said repeatedly, in one form or another, after the election. When fuzzy interpretation was put up against statistical analysis, the stats won out.

But not every journalistic question benefits from the data set that Silver and the other electoral quants had to work with. What if you want to measure a question that doesn’t offer the cleanly defined Obama-or-Romney options of an election? Or what if you’re interested in the opinions of a group that isn’t defined by the electoral boundaries — or one where there just isn’t good polling available to aggregate?

These are the sorts of questions where journalism turns to pundit representatives. Want to know what “hockey moms” think? Sarah Palin gets called upon to represent them. How about the Occupy movement? Call Michael Moore. The Tea Party? Bring in Dick Armey. Gun owners? Alert Alex Jones. This sort of representative punditry comes with obvious, distorting flaws: Alex Jones doesn’t represent all gun owners and Michael Moore doesn’t represent everyone on the activist left, but the workflows of contemporary journalism let both stand in for what a larger group is thinking or feeling. And if your group doesn’t have an obvious mouthpiece, someone already in a cable news producer’s contacts? You might just get excluded from the narrative altogether.

That’s one reason news organizations are increasingly turning to sentiment analysis to try to reflect a crowd’s thoughts through data instead of punditry. When used wisely, sentiment analysis can push back against sweeping punditry claims and let many voices contribute to a message.

Sentiment analysis aims to analyze what a percentage of the population “feels” about something, often by measuring the sentiments embedded in social media posts or by asking a community directly to share its feelings, thoughts, or opinions in a machine-readable way. Many news outlets have already deployed sentiment analysis, to various effects. Politico, Pew, NBC, CNN, Current TV, Twitter, Facebook, and many others have all tried to measure the thoughts and feelings of social media groups or run these analyses a regular basis. For example, after the GOP primary, an analysis by Politico showed the amount of negativity expressed towards each the candidates on Facebook during the most intense months of the primary season. Newt Gingrich led handsomely.

The use of machines — or machine learning — in this way to discern the meaning of human feeling about a topic is by no means unique to social media. The New York Times wrote about “affective programming,” and how computers may soon read your facial expressions and respond to you. Another Times piece famously explained how Target can now determine when you are pregnant based on your consumer habits — sometimes, before you realize yourself.

What’s interesting about the use of sentiment analysis by journalists, though, is that so many of the industry’s ongoing concerns seem to crystallize in its promise: how to deal with social media platforms where the power to publish now belongs to millions; how to find a way to speak more authoritatively about the world it reports on; and how to take complex questions and display them simply and visually.

And it’s true that sentiment analysis does offer real potential on these question. Already, the technology exists to fire up software to go into a group Facebook page, run through specific posts, and gather a list of most discussed themes and general positive or negative reaction to specific debate points. This could be turned onto, say, The Tea Party Patriots page, with nearly 1 million users, to enrich knowledge of “the Tea Party opinion” on an issue. And at times, sentiment analysis has proven to be smarter than other data sources, as when it predicted Harry Reid’s 2010 reelection campaign was looking stronger than polls suggested.

Bu how accurate is this stuff? As Micah Sifry, co-founder of the Personal Democracy Project, said to me: “Whenever you hear the words ‘sentiment analysis,’ your BS detector should go up.” Sifry wrote a bearish take on Politico’s use of the technology, calling for them to stop using it and explaining some of the internal problems with machine learning.

Without going too deep into the weeds, natural language processing’s problems can be imagined by thinking about the “sentiment” of the following statements:

— “It is a great movie if you have the taste and sensibilities of a five-year-old boy.”
— “There’s a lot of tumult in the Middle East.”
— “It’s terrible Candidate X did so well in the debate last night.”

These examples — provided by Christopher Potts, a linguistics professor at Stanford who teaches natural language processing for the school’s computer science department — show the complexity. “Tumult” might sound like a negative sentiment, but if it’s in the context of a democratic revolution the speaker supports, maybe it’s not. In the first quote, it’s hard to know the speaker’s goals: Is this someone speaking tongue-in-cheek as a mom to other moms? As Potts says: “If you and I are best friends, then my graceful swearing at you is different than if it’s at my boss.” There is a certain threshold of knowing what other people mean when they speak that is still hard for machines to reach without additional information.

One popular tool, Crimson Hexagon — which has been used by CNN, NBC, Pew, and Current TV — claims its technology produces 97 percent accuracy. But what does “accuracy” mean in this context? It doesn’t mean Crimson Hexagon goes one-by-one to the authors of the analyzed posts and asks them what they really meant and that they’re right 97 times out of 100. Instead, it measures the tool’s accuracy relative to a hypothetical group of humans — a hypothetical test group who read the same tweets or Facebook posts and gauged what their authors were thinking. Crimson Hexagon tests its product against human groups in this way. So its accuracy is basically a rate at which it keeps up with “the crowd.”

So how accurate is the crowd? On this, too, researchers disagree, because there is no single interpretation of language. Anywhere from 65 percent to 85 percent is a reasonable guess, depending on the difficulty of the analysis. In this video from the Sentiment Analysis Symposium, Shawn Rutledge, chief scientist at Visible Technologies, discusses one sentiment-analysis study his company performed for a financial institution. The good news: The machine-driven analysis was found to be very close to an analysis by humans, with no statistically significant difference between the two. The bad news: “The problem is that the reviewers thought that the humans sucked at annotating sentiment just about as much as the algorithms.” (Philip Resnik, who teaches linguistics and computer science at the University of Maryland, says one possible way to label the results’ accuracy would be as a “percentage of human performance” — consistent with the idea that all analyses are judged against a test-control group, not objective reality.)

Setting a machine algorithm on hundreds of thousands of tweets about Mitt Romney’s statement about “binders full of women” and asking it to parse the thick cloud of sarcasm and irony, for example, is a tough nut to crack. Software can hit snags when trying to grasp complex language, or it can encounter bots or nonsense that isn’t filtered out. When used carelessly, sentiment analysis’ results can be absurd.

And that’s all before considering the difficulty of using any social media universe to represent a broader group of people. During the election, many polls were criticized for methodologies that undercount voters who had only mobile phones, noting that they likely undercounted young and less-wealthy voters. A social-media-centric view of any population likely brings its own distortions — for example, possibly undercounting the elderly or other groups who might be less likely to livetweet a presidential debate.

For news organizations, then, the question is whether this imperfect accuracy should keep the tool from being used. I don’t think so: All sampling technologies have problems, including the very best polls news organizations produce or write about. What’s essential is how results are presented — with the appropriate context and caveats about its limits.

Someone who has made headway at that is Alex Johnson at NBC News, who for the last year performed a number of Crimson Hexagon-driven sentiment analyses. Johnson wanted to find a way to try to say something about people’s online conversations and had some background analyzing data, so he spent time reading through a grip of white papers, statistical models, and documentation. He finds the technology “still pretty rough,” but thinks it adds value to NBC.

Johnson’s first problem was how to display results while being transparent about its inherent problems and setting it apart from normal polling. “The number one rule we have is we want to make it crystal clear that this is not a poll,” he says.

He experimented with a number of disclaimers through the election cycle. At some points, he linked to and tweeted a blog post he in which he describes his methodology. He tries to explain upfront that this is a new technology that only captures a segment of the population and he stays away from saying it is accurate to a specific percentage level. His guiding philosophy, which he picked up as a newspaper reporter, is that your piece should support your lede, which he considers in this case the analysis results. “Let people see how you did it and give people a basis to challenge you,” he says.

Some of Johnson’s most revealing findings are in trends in the conversations, which create linguistic maps around concepts or ideas.

Johnson’s brief analysis from this graphic: “A representation of key words in comments that said Ryan did better illustrates the degree to which his performance was defined in relation to Biden’s. Notice that the word ‘Biden’ is fully as prominent as the word ‘Ryan.'” Reporters could present this result as in keeping with an influential body of political science research that finds conservatives tend to define themselves “against” liberals.

Johnson also recommends running analyses over a long time-window — like trawl-fishing over a massive ocean. Long after an event, he finds fluctuations in feeling at pressure points, such as the Sunday talk shows, specific media ad buys, and other follow-up events.

Johnson also emphasizes that this technology only reads a specific userbase — those on social media. But even within that limited universe, a group-based analysis can be a powerful way to challenge those people representing — or claiming to represent — those groups. The key is to not let imperfect sentiment analysis take charge of an overall coverage strategy: It’s a conversation starter, not a conversation ender. As Simon Rogers, editor of the Guardian Data Blog — which worked across the Guardian newsroom to do a highly successful sentiment analysis of tweets during the London riots — says: “It’s when you combine traditional journalistic know-how with a manual check that makes something much more rigorous and trustable.”

Whether using existing sentiment analysis software — Crimson Hexagon, Topsy, Radian6 —or building your own tools with open-source software such as R and code from GitHub, the potential for new forms of human feedback exists. Nate Silver puts this idea succinctly in his book: “Data-driven predictions can succeed — and they can fail. It is when we deny our role in the process that the odds of failure rise.”

Much reporting about activity on social media end up being purely quantitative — think of the regular reports on how many tweets per second were generated at each of the election’s high points. As Lydia DePillis put it in TNR: “Really cutting-edge Twitter analysis means that interested parties can talk back.” When media outlets engage Twitter conversations, they encourage this participation to flourish, and even quirky tools can provide a placeholder for kicking things off. The media already does plenty of “sentiment analysis” — every time a commentator divines what voters do or do not really “care about,” or what Americans are or not “really thinking” about at any moment, as if this can be easily discerned with a quick read from the tea leaves. As Nate Silver showed, the gut, often self-interested, is far from foolproof. Even though it still has a lot of room for improvement, machine-learning sentiment analysis can, when used properly, bring more people into the conversation.

POSTED     Jan. 23, 2013, 10 a.m.
Show tags
Join the 60,000 who get the freshest future-of-journalism news in our daily email.
Postcards and laundromat visits: The Texas Tribune audience team experiments with IRL distribution
As social platforms falter for news, a number of nonprofit outlets are rethinking distribution for impact and in-person engagement.
Radio Ambulante launches its own record label as a home for its podcast’s original music
“So much of podcast music is background, feels like filler sometimes, but with our composers, it never is.”
How uncritical news coverage feeds the AI hype machine
“The coverage tends to be led by industry sources and often takes claims about what the technology can and can’t do, and might be able to do in the future, at face value in ways that contribute to the hype cycle.”