You have the full HN dataset and voting logs. What if you used it to train a simple classifier that would predict the likely karma of a comment as a function of a number of features, including username, username responded to, bag-of-words vector of the comment, time of day, total number of comments in thread, "controversy" of the comment (as reflected by a mix of up & down votes) and similar factors?

The goal could be something like Google's Beer Goggles, but for comments: "it looks like you're writing a flame, are you sure you want to do that?"

Another thought: there's a big difference between a comment at -1 with 10 upvotes and 12 downvotes vs. one that has occasioned 2 downvotes. Might be interesting to see the effect of showing both the number of up & down votes rather than just the sum. The idea is that genuinely trollish comments at -1 will be differentiated from minority points of view.