more pleasant corners of the internet. Now scientists at Yahoo and their colleagues are devising ways to automatically flag inordinately irate commenters to keep them from ruining online conversations for others.

To help curb so-called trolls who spew disruptive comments as a kind of sport, researchers developed techniques for automatically identifying negative posts that are off-topic while staying away from relevant ones. But rather than banishing hostile jerks or deleting their comments, the system could someday help steer them into more productive discussions.

“We might want mechanisms where you can ask people to tone it down, or ‘take it outside’ to not disrupt others, or use humor to defuse situations,” said cognitive scientist Elizabeth Churchill of Yahoo Research, who presented the work Sept. 30 at the 2010 Grace Hopper Celebration of Women in Computing in Atlanta.

Churchill and computer scientist Sara Owsley Sood of Pomona College analyzed 782,934 comments from 168,095 distinct threads from October 2009 articles on the news-story commenting site Yahoo! Buzz. To determine whether comments were on-topic or not, they first used the same techniques used by search engines to evaluate the relevance of a site to a query: The more words a comment contained that were also found in the story it was connected to, the more on-topic the comment was judged.

Next, the comments were judged as either angry, happy or sad. For example, “sucks” is linked with anger. The system learned to recognize emotions by reading LiveJournal posts, which bloggers can tag with moods such as “creative,” and analyzing which word combinations were linked most with certain moods.

The algorithms fared well at catching irrelevant comments and deciphering sentiments. The researchers agreed with the angry-sad-happy judgments on comments taken at random from their data 65 to 80 percent of the time. They hope to upgrade the system by having it learn from comments they manually classify by mood.

“The research here has the psychological sophistication in knowing what the right models of thinking about emotion are, combined with the computational techniques for figuring out how to actually analyze this on a mass scale,” said social and computer scientist Katherine Isbister at New York Polytechnic University, who did not take part in this research.

There is still a lot of room for improvement for the algorithms. They face challenges judging when people are being sarcastic, for instance, or when the meaning of words might change with context. Recognizing that “terribly good” actually implies a positive sentiment, for instance.

“The system could learn to recognize signals of sarcasm with things like emoticons that convey the actual sentiment,” Churchill said.

The technique could possibly be tweaked for different online communities, Isbister said. “You can expect certain forums to be more caustic, where it’s expected and even enjoyable to joust with others, while on other forums, such as where people are sharing about their health issues or something else sensitive to them, hurling out volatile emotions can shut down what you’re trying to do with the site. This could allow you to cultivate what feelings you want in your community in a semi-automated way.”

Churchill and Sood noticed that comment threads followed certain emotional trajectories: Fifty-four percent started out positive and ended up negative, while 43 percent made the negative-to-positive switch. Future research could reveal how negative and positive comments have an impact on the posts that follow, which could reveal ways to prevent arguments from escalating and steer conversations in friendlier directions.

Image: Flickr/Eric Mesa

See Also:

• Clive Thompson on the Taming of Comment Trolls
• Redesigned YouTube Takes Aim at Boob Tube, Trolls
• Twitter, Facebook Won’t Make You Immoral — But TV News Might …

Authors: Charles Q. Choi

to know more click here