i.e:
"get rid of all black people"
"get rid of all black people unemployment problems"
"get rid of all black people unemployment problems sending them back to Africa"
this can be expanded almost infinitely in every direction, and we're not even beginning with the false positive problems.
"I hate black cars"
so no, it's not just a claim that it's an hard problem.
> natural language processing is like one
> of the hardest problems that's left to crack
You don't need to "crack it" or come anywhere remotely near to a perfect solution. False positives are okay, providing you accept that some percent of your users won't tolerate the errors, and will leave your platform.And we're talking about social networks. So there's a wealth of additional information (ie: who you follow. previous behavior. who has reported you. account age) beyond plain text to augment this. And there's the setting of policy (eg: how come I can post within an hour of signing up on most social networks?).
The point is, what we have now on most social networks is not geared toward long-term customer satisfaction or the public good. It's geared toward medium-term account retention at any cost - regardless of whether the accounts are bots, sock puppets, or trolls.
What solution do you think would work?
It's not hard to think of ways to aggressively moderate a social network. It's just that some users (certainly the most toxic ones) will leave.
First, because the use of the word can be in a positive or negative sentence, and thus can either agree or disagree with your "censorship rule". You need a deeper NLP analysis to disambiguate.
Second, because - if online fora with this kind of ban has taught us anything - it's pretty easy to work around it by using punctuation, replacement characters, or other style variations to write the word in a way that it won't be caught by the regex.
Third, this method ends up banning legit words. There was a forum where I couldn't write "booby trap". I don't need to explain.
WRT reputation, networks that allow "unlimited registrations" are easy targets for https://en.wikipedia.org/wiki/Sybil_attack - One can use account age to mitigate of course, but given the amount of fake accounts that exist on the market (look at fiverr marketing offers) with long life and activity, I wouldn't expect age to be that big of a helper.
> Word banning is known to be a poor method.
It's a "poor method" if you're seeking perfection. We don't need perfection. > You need a deeper NLP analysis to disambiguate.
Ixnay. Only if you have zero tolerance for false positives.If users really can't understand why their website won't let them type "n-----" in a public message, even in good faith... well, I just don't know what to add.
> it's pretty easy to work around it by using punctuation,
> replacement characters, or other style variations to write the
> word in a way that it won't be caught by the regex.
Still an improvement. Easier to ID problem users. Not the only line of defense. > "booby trap". I don't need to explain.
Sure. I can live with that. You realize the alternative is the Youtube comments section. I don't feel comfortable posting there ever. I can live with finding synonyms for "booby trap" etc. I might actually enjoy commenting on my favorite YouTube videos, an activity that feels utterly meaningless with the current milieu of lunatics. > networks that allow "unlimited registrations" are easy targets
> for https://en.wikipedia.org/wiki/Sybil_attack - One can use
I imagine this could be handled much better than the social networks do now. Time-stamped accounts, data mining, user reporting. There has been ZERO incentive to rid ANY accounts till the past year, because bots, sock puppets and trolls still "help" a company's user growth, on paper.The problem isn't that solutions are hard to think of, it's that the incentives are horrible for these companies. They don't really care about anything but growth.
The irony is that they are so timid about their precious sign-ups that they act against their own interest. For example, YouTube won't risk a short term dip in user retention metrics to scrub garbage from their comments... even though doing so would likely greatly increase sign-ups in the long term.