Here's a wild guess: train a model to predict average comment karma and replies given a comment's depth in a given thread. The select the top 3 comments, plus

Here's a wild guess: train a model to predict average comment karma and replies given a comment's depth in a given thread. The select the top 3 comments, plus any sub-comment with a combination of karma and replies higher than 1 or 2 standard deviations. You could do some fancy text analysis, but seeing as you already have humans doing that (which is where karma comes from), I doubt you could improve that much.