Earlier quoted context omitted.
The problem is that word features are not really that good predictors of quality. I have done some research on this (unpublished), and I got a really good performance on predicting hacker news votes by just counting how many new words (not stopwords, not very-high-frequency words) a comment was adding to a thread. Just using a few variations on this theme predicted better than word counts or bigram features. Fundamen…
alextp, what if you use ML on voting patterns instead of comment words? (also what if the ML only provided feedback while one is typing the comment?)
Using ML to provide feedback is a bad idea. Most ML techniques latch on to surface features of the text rather than the deeper structure, so it'd just make it really easy for people to reword their mean comments ("this is just stupid" becomes "What an incoherent piece of gobblydegook" or something like this, which might make things funnier but I doubt it would help).