Yeah I’m curious if they tried training a Bert or similar classifier… intuitively this seems better than tfidf which is throwing away a ton of information.