Yahoo's New Anti-Abuse AI Outperforms Previous AI (wired.co.uk)

16.4% of the comments on Yahoo News are "abusive," according to human screeners. Now Yahoo has devised an abuse-detecting algorithm "that can accurately identify whether online comments contain hate speech or not," reports Wired UK: In 90 per cent of test cases Yahoo's algorithm was able to correctly identify that a comment was abusive... The company used a combination of machine learning and crowdsourced abuse detection to create an algorithm that trawled the comment sections of Yahoo News and Finance to sniff out abuse. As part of its project, Yahoo will be releasing the first publicly available curated database of online hate speech.
The machine-learning algorithm was "trained on a million Yahoo article comments," according to the article, and Slashdot reader AmiMoJo writes "The system could help AIs avoid being tricked into making abusive comments themselves, as Microsoft's Tay twitter bot did earlier this year."

Comments

"Hate speech"

[axewolf]

Daily reminder that if you subscribe to this idea of "hate speech" you are totally insane and have no mind of your own.

Speech can be hateful, leave it at that, everyone knows it well, no need to make a special idea for it that can also be manipulated to cover valuable critical thought.

Re:"Hate speech"

[iggymanz]

yahoo news and abcnews hate dissenting non-SJW type viewpoints and label them "abuse"

Re:"Hate speech"

[NotQuiteReal]

They have to label it something other than "thought crime", because people aren't ready for that. Yet.

Re:"Hate speech"

[UnknownSoldier]

Exactly. Hate Speech == Censored Speech.

It's fine to have your own opinion but when you dictates that others follow your own myopic viewpoint because you're too insecure, congratulations, you've just resorted to censorship.

Ignoring something doesn't make it go away.

--
SJW, noun, acronym for Stupid Justice Whiner

Re:"Hate speech"

[RobotRunAmok]

Understand that modding "redundant" is the Slashdot SJW's way of saying, "Whuuut? How could so many people on Slashdot think differently from me? I'm too afraid to make an argument against them, but maybe if I mod 'redundant' I will scare off anyone else who wants to chime in with a point of view that makes me feel sad."

Re:"Hate speech"

[AmiMoJo]

What phrase would you use to describe hateful speech that commercial AIs may wish to avoid using? Keep in mind the target market for these bots is customer service and advertising, so avoiding going full "Tay" with "Hitler did nothing wrong!" is a design goal.

Not everything is a conspiracy against free speech.

Re:"Hate speech"

[Dutch Gun]

From the article:

The comment dataset came from Yahoo! Finance and News, which you wouldn’t think of as exactly the dank basement of the internet—but it turns out a whopping 7% of comments on Finance and 16.4% on News were deemed abusive by human screeners.

I look forward to perusing this database and finding out exactly how abusive these comments actually are. I don't doubt there are a lot of assholes on the internet. Slashdot has one of the most effective self-policing filtering mechanisms I've seen. Browse at -1 on occasion to see some of the crap, even though I think most obvious trolls have left, as they can't get much traction here. Sure, it goes awry sometimes during heated debates (toss the word SJW in there and you're going to spark some positive and negative mods just with that), but overall does a good job of filtering out most of the garbage. But what's the criteria? Are we talking genuine abuse, or "microagressions", where someone perhaps just expresses an opinion that's not quite PC enough for someone else's delicate ears?

Moreover, a 90% success rate doesn't strike me as all that fantastic. That means 1 in 10 abuse flags is a complete false positive. That would be a completely unacceptable rate with a spam filter. I have a feeling that with all the training that's been done, all you have to do is say one of the "banned" words, like "cunt", and you'll get flagged, even if speaking about it in an appropriate context, like I'm doing. Or am I being abusive just by mentioning the word, since I'm a man? Oops, probably another flag there for mentioning my male gender. Given the tone of my post just now, I'll bet even a human might flag me for a microagression.

Unlike some here, I think the effort to civilize online speech is not unwarranted. Talk to prominent men and (especially) women online and find out how much verbal abuse gets heaped on them. It's fairly disgusting. But this is an area which, I think, needs to be tread upon very carefully. Simply blacklisting hateful speech doesn't cure the problem. It addresses a symptom, and will just push the abuse "underground" to a level where an AI can't detect it. Humans are clever that way.

I'm not sure what can be done short of preventing anonymous interaction, because that seems to bring out the worst in people. Many notorious trolls tend to be cowards, and are mortified when their actions are associated with their real names. This algorithm is obviously one way to address the problem, but I have to admit I remain skeptical about this sort of approach.

And so it has begun...

[MindPrison]

...get ready for the ultimate PC society:

ModBot: I am sorry, but your message was removed because of Violation of rule #157792 - negative opinion on political minority group.

ModBot: I am sorry, but your message was removed because of Violation of rule #151734 - negative opinion of product. Be fruitful!
ModBot: I am sorry, but your message was removed because of Violation of rule #191727 - hate speech: you voiced an opinion on criminals. Let's leave that to our law enforcement, right? Be well!

ModBot: I am sorry, but your message was removed because of Violation of rule #1 - Personal Opinion. We encourage our citizens to support each other, personal opinions are best kept to yourself, be well citizen!