net.wars: The sentiment of crowds
by Wendy M Grossman | posted on 11 November 2011
Context is king.
Say to a human, "I'll meet you at the place near the thing where we went that time," and they'll show up at the right place. That's from the 1987 movieBroadcast News: Aaron (Albert Brooks) says it; cut to Jane (Holly Hunter), awaiting him at a table.
But what if Jane were a computer and what she wanted to know from Aaron's statement was not where to meet but how Aaron felt about it? This is the challenge facing sentiment analysis.
At Wednesday's Sentiment Analysis Symposium, the key question of context came up over and over again as the biggest challenge to the industry of people who claim that they can turn Tweets, blog postings, news stories, and other mass data sources into intelligence.
So context: Jane can parse "the place", "the thing", and "that time" because she has expert knowledge of her past with Aaron. It's an extreme example, but all human writing makes assumptions about the knowledge and understanding of the reader. Humans even use those assumptions to implement privacy in a public setting: Stephen Fry could retweet Aaron's words and still only Jane would find the cafe. If Jane is a large organization seeking to understand what people are saying about it and Aaron is 6 million people posting on Twitter, Tom can use sentiment analyzer tools to give a numerical answer. And numbers always inspire confidence…
My first encounter with sentiment analysis was this summer during Young Rewired State, when a team wanted to create a mood map of the UK comparing geolocated tweets to indices of multiple deprivation. This third annual symposium shows that here is a rapidly engorging industry, part PR, part image consultancy, and part artificial intelligence research project.
I was drawn to it out of curiosity, but also because it all sounds slightly sinister. What do sentiment analyzers understand when I say an airline lounge at Heathrow Terminal 4 "brings out my inner Sheldon? What is at stake is not precise meaning – humans argue over the exact meaning of even the greatest communicators – but extracting good-enough meaning from high-volume data streams written by millions of not-monkeys
What could possibly go wrong? This was one of the day's most interesting questions, posed by the consultant Meta Brown to representatives of the Red Cross, the polling organization Harris Interactive, and Paypal. Failure to consider the data sources and the industry you're in, said the Red Cross's Banafsheh Ghassemi. Her example was the period just after Hurricane Irene, when analyzing social media sentiment would find it negative. "It took everyday disaster language as negative," she said. In addition, because the Red Cross's constituency is primarily older, social media are less indicative than emails and call center records. For many organizations, she added, social media tend to skew negative
Earlier this year, Harris Interactive's Carol Haney, who has had to kill projects when they failed to produce sufficiently accurate results for the client, told a conference, "Sentiment analysis is the snake oil of 2011." Now, she said, "I believe it's still true to some extent. The customer has a commercial need for a dial pointing at a number – but that's not really what's being delivered. Over time you can see trends and significant change in sentiment, and when that happens I feel we're returning value to a customer because it's not something they received before and it's directionally accurate and giving information." But very small changes over short time scales are an unreliable basis for making decisions.
"The difficulty in social media analytics is you need a good idea of the questions you're asking to get good results," says Shlomo Argamon, whose research work seems to raise more questions than answers. Look at companies that claim to measure influence. "What is influence? How do you know you're measuring that or to what it correlates in the real world?" he asks. Even the notion that you can classify texts into positive and negative is a "huge simplifying assumption".
Argamon has been working on technology to discern from written text the gender and age – and perhaps other characteristics – of the author, a joint effort with his former PhD student Ken Bloom. When he says this, I immediately want to test him with obscure texts.
Is this stuff more or less creepy than online behavioral advertising? Han-Sheong Lai explained that Paypal uses sentiment analysis to try to glean the exact level of frustration of the company's biggest clients when they threaten to close their accounts. How serious are they? How much effort should the company put into dissuading them? Meanwhile Verint's job is to analyze those "This call may be recorded" calls. Verint's tools turn speech to text, and create color voiceprint maps showing the emotional high points. Click and hear the anger.
"Technology alone is not the solution," said Philip Resnik, summing up the state of the art. But, "It supports human insight in ways that were not previously possible." His talk made me ask: if humans obfuscate their data – for example, by turning off geolocation – will this industry respond by finding ways to put it all back again so the data will be more useful?
"It will be an arms race," he agrees. "Like spam."
Technorati tags: sentiment analysis social media twitter
Wendy M. Grossman’s Web site has an extensive archive of her books, articles, and music, and an archive of all the earlier columns in this series. Readers are welcome to post here, at net.wars home, follow on Twitter or send email to netwars(at) skeptic.demon.co.uk (but please turn off HTML).