I just read Paul's essay in Hackers and Painters on his spam filters using Bayesian probability of words in the message. Could something like that work on trolls? The body's corpus would be only replies marked with a certain flag?