Were you able to get a bayesian algorithm to learn whether specific users were likely to be spammers or trolls? (As opposed to specific posts.) I'm guessing this isn't easy to do, since I get spammed on other services on a regular basis (e.g. Twitter spammers will setup new accounts and spam on a regular basis).
i believe it's possible to run a node and just query it for the exit nodes, but I'm not sure.
I would, in fact, be somewhat surprised if nobody has proposed deploying this kind of thing against, say, Anonymous.
Say you use an n-gram analysis at the word level. I can just build a tool that processes a corpus, and then as I type, suggests the most probable next word for me. I can simply accept those words when appropriate and look like a different person.
From the article, we have a troll writing "Rest in piss" and "Rest in piss, you n*r". I'm pretty sure those won't personally identify anyone.
Only very rarely do I get traceable IP addresses that can be dealt with in any meaningful way, and that's usually when there is an invasion of trolls for a day or two and some drop their guard.
The only time I've ever successfully stopped a troll was about a decade ago when I traced it back to a uni and was able to raise a sysadmin.
By and large, for the majority of forums and social sites out there (which is where most bullying happens) there is very little that the admins can do even if it's reported to us.
About the only thing working is stopforumspam.com , whereby hundreds of forums are blocking IP addresses for a short while across all of the forums. It's made for spam, but I and others submit trolls to it too (when we're absolutely sure there's nothing of worth in the person and it's not just differing views).