3 ms·
Reposting this paper yet again, to rub in the point that social media platforms play host to communities and communities are often very good at detecting interl
by anigbrowl 4y ago
Reposting this paper yet again, to rub in the point that social media platforms play host to communities and communities are often very good at detecting interlopers and saboteurs and pushing them back out. And it turns out the most effective approach is to let people give bad actors a hard time. Moderation policies that require everyone to adhere to high standards of politeness in all circumstances are trying to reproduce the dynamics of kindergartens, and are not effective because the moderators are easily gamed.
https://arxiv.org/pdf/1803.03697.pdf https://arxiv.org/pdf/1803.03697.pdf
Also, if you're running or working for a platform and dealing with insurgencies, you will lose if you try to build any kind of policy around content analysis. Automated context analysis is generally crap because of semantic overloading (irony, satire, contextual humor), and manual context analysis is labor-intensive and immiserating, to the point that larger platforms like Facebook are legitimately accused of abusing their moderation staff by paying them peanuts to wade through toxic sludge and then dumping them as soon as they complain or ask for any kind of support from HR.
To get anywhere you need to look at patterns of behavior and to scale you need to do feature/motif detection on dynamic systems rather than static relationships like friend/follower selections. However, this kind of approach is fundamentally at odds with many platforms' goal of maximizing engagement as means to the end of selling ad space.