would probably work better with case and punctuation preserving n-grams, sentence length, paragraph length and use of whitespace stats. also maybe a tf-idf vec

would probably work better with case and punctuation preserving n-grams, sentence length, paragraph length and use of whitespace stats. also maybe a tf-idf vector of top n words per user. also could maybe do a same phrase analysis across the corpus to find some hand picked features. timestamps could be interesting. or, of course, let the machine do it with comment2vec.