Sister blog of Physicists of the Caribbean in which I babble about non-astronomy stuff, because everyone needs a hobby

Sunday 15 October 2017

An AI that can distinguish hate speech from harmless banter, based on context

Instead of focusing on isolated words and phrases, they taught machine learning software to spot hate speech by learning how members of hateful communities speak. They trained their system on a data dump that contains most of the posts made to Reddit between 2006 and 2016. They focused on three groups who are often the target of abuse: African Americans, overweight people and women. For each of these, they chose the most active support and abuse groups on Reddit to train their software. They also took comments from Voat -- a forum site similar to Reddit -- as well as individual websites dedicated to hate speech.

The team found that their approach contained fewer false positives than a keyword-based detector. For example, it was able to flag comments that contained no offensive keyword, such as 'I don't see the problem here. Animals attack other animals all the time,' in which the term 'animals' was being used as a racist slur.

Interesting. But can it differentiate between hatred and momentary outrage ? What about hate that's been provoked and/or is justified, i.e. hating Nazis ? What about in countries where violent political action might actually be required, e.g. those under the rule of a brutal dictator ?

I'm no free speech absolutist - totally free speech is a fundamental impossibility - but only a fool would pretend that the situation isn't horribly messy.

https://www.newscientist.com/article/2149562-this-ai-can-tell-true-hate-speech-from-harmless-banter/

4 comments:

  1. Here's where I think this is going: Facebook have thousands of Indians and Filipinos on the job, doing content moderation. It's serious business already.

    wired.com - The Laborers Who Keep Dick Pics and Beheadings Out of Your Facebook Feed | WIRED

    Betteridge's Law of Headlines here: all the situation needs is a better gold sluice box. Just filter off the non-problematic posts and comments, the humans can handle the rest

    ReplyDelete
  2. There's always a danger of human bias in creating the training dataset in situations like this.

    ReplyDelete
  3. Being firm is NOT being rude (its telling you what others are afraid of telling you straight away)
    #carbonemissions
    #climatechange 2day in South Africa :;

    ReplyDelete

Due to a small but consistent influx of spam, comments will now be checked before publishing. Only egregious spam/illegal/racist crap will be disapproved, everything else will be published.

These things are not the same as these other things

Today, a couple of similar-ish pieces from Pscyhe I think I can get away with combining into a single post. The first one is very simple, d...