zlacker

Bayesian filtering for comments based on downvotes, with failing comments requiring moderation by a human. My guess is that most mean/dumb/ad-hominem comments use the same words/phrases as other mean/dumb/ad-hominem comments.

To prevent moderation becoming a bane on non-spamming users, the threshold could be set to .05 or .10. (or .95/.90, depending on which way you train the filter.)

It worked for email when spam was a massive problem; perhaps it can work for comments while they're still only s small problem.