While I was reading a forum a few minutes ago, I got an idea of how to do moderation when there aren't enough moderators to moderate all the messages every time.
Consider cranks, for instance Usenet kooks. These people tend to have a very particular way of writing, and they usually also try to pull the conversation away from whatever subject is being discussed into their pet subjects or peeves. The combination of these two traits should make it very simple to detect, when a post has been written, whether it's written by a crank, even if that crank tries to hide himself by using proxies, sock puppets, etc.
So here's an idea: "Bayesian moderation". This is like spam filtering, only with moderation instead. It's somewhat of a misnomer since it doesn't have to use naive Bayes models, but that name has stuck for spam filters, so why not use it here, too?
Bayesian moderation would work like this: Before the model is trained, moderation is done after the fact. Anything that is removed also updates the classification model, so the model gets better at predicting what would be removed and what wouldn't be removed. When the model is properly trained, it holds posts that fall above a certain threshold of crankiness (that is, probability that were a moderator to see it, he would delete the post). Actual moderators examine each of the held posts and decide either to remove the post or let the post be published; and depending on the decision, the classification model is either updated to consider the text less cranky (if the post was published) or more cranky (if it was deleted).
With a proper model, the system could also rank posts in either order of crankiness or uncertainty (if the model calculates uncertainty), so that the limited time the moderators have would be used on the posts that most likely need to be investigated.
![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}