Blowfish12: AI systems could fight cyberbullying

Wednesday, July 4, 2012

AI systems could fight cyberbullying

Smart software could detect online bullying via a database system that can identify even the subtlest of abusive comments

"I have been bullied my entire life. About how I look like a whale and how im not pretty enough. I cant get boyfriends because i refuse to have sex until I am married. I just dont know what to do anymore...:\" - Samantha, 16

Pleas for help like this one appear on social media and internet forums every day, written by desperate teenagers who live their entire lives online. Knowing you're not alone can help. That's the idea behind new software that matches up such messages with similar posts from other worried teenagers, letting them know that what they're experiencing isn't unusual. It might also be possible to spot bullying behaviour as it happens online.

Recent high-profile cases have made cyberbullying front page news. In January, 15-year-old Amanda Diane Cummings died after jumping in front of a bus on Staten Island, New York. She'd been subjected to a campaign of bullying on Facebook by other pupils at her school. Last September, Jamey Rodemeyer, a 15-year-old boy from Buffalo, New York, killed himself after being teased online about his sexuality. The cases sparked lawmakers to push through legislation, passed by the state senate last week, that makes cyberbullying a crime.

To help tackle one part of the problem, Karthik Dinakar at the Massachusetts Institute of Technology and colleagues have been working on a project that analyses the posts written by teenagers on A Thin Line, a website run by MTV. The site encourages teenagers to post their problems anonymously and other teenagers leave comments giving advice. Many of the posts concern bullying and worries about sex.

Each of the website's 5500 posts were fed through an algorithm trained to recognize certain clusters of words and then categorize each post according to one or more of 30 themes, ranging from "duration of a relationship" to "using naked pictures of girlfriend". The words "boyf" "trust" "cheat" "break" "upset" in the same story might indicate the post was about a relationship ending, for example. Once a label was assigned, the algorithm picked another story on the site that covered the same themes.

"All these teenagers are still growing emotionally, and there's a tendency to think that their experience is singular to themselves," says Dinakar. "It can let them know that they are not alone in their plight."

The software was tested using a set of new stories written by volunteers, which it analyzed and matched with stories from the website. The volunteers rated the system very positively. They felt that the stories picked using the thematic algorithm were always a much closer match than those chosen using a basic algorithm that just matched keywords. The system was presented at a conference on social media in Dublin, Ireland, earlier this month. MTV now plans to start using it to match stories live on the site, so teenagers can read about those in a similar plight.

Can artificial intelligence also stop cyberbullying at its source? After Amanda Cummings died, her memorial Facebook page was filled with offensive comments, leaving her parents understandably distraught. So Dinakar is also developing software that will help spot online bullying as it happens.

Facebook has taken steps to stop cyberbullying, but it primarily relies on users flagging up comments as inappropriate.

To find less-obvious forms of abuse, Dinakar built software that compares online posts to an open-source database called ConceptNet. This is a network of phrases and words and the relationships between them that lets computers understand what humans are talking about. This way the system can work out what might be a bullying comment, even though it contains no abusive words. For example, it would know that: "Put on a wig and lipstick and be who you really are" aimed at a boy might be a negative comment on his sexuality, because ConceptNet knows that girls usually wear make-up, while boys do not.

The idea is that software like this could be integrated into a social network. If it spots patterns of bullying behavior, it may either flash up a box warning the bully, ban offending posts, or offer help and advice to the victim. Dinakar wants to combine his two projects to create a detector that can pick up even the subtlest of attacks, such as "liking" a negative Facebook status to make a nasty point, for example. The research is due to appear in the journal ACM Transactions on Interactive Intelligent Systemsin July.

Danah Boydof Microsoft Research in Cambridge, Massachusetts, says that although this kind of work won't solve the problem of online bullying, it will help to improve our understanding of what happens online.

"I'm glad that these researchers are working to identify different types of meanness and cruelty," she says. "I am very hopeful that these kinds of techniques will lead to a more holistic understanding of the problem."

Blowfish12@2012 blowfish12.tk Author: Sudharsun. P. R.