> If anyone could suggest ways of distinguishing truly flamey threads from those which are substantive-but-spicey, I'd love to hear them
Taking threads that trigger the flamewar detector (simple comments vs points comparison) and then running them (either individual comments or the comments as a whole) through an LLM to score them for flamebait
seems obvious so my question is why didn't that go anywhere. If you didn't want to use new technology, you could still just build a classifier based on the usernames involved. It wouldn't have to be perfectly accurate, just enough to flag something for manual review.
There's a simpler heuristic which was developed at Microsoft back in the early aughts, based on Usenet, which simply looked at thread topology to make inferences about the type of exchange.
I'm pretty sure that was Marc Smith (the name was a very common one, that or Jones was in my memory), see:
"You Are Who You Talk To: Detecting Roles in Usenet Newsgroups" (~2005?)
(Date based on citations and recent data (2004) mentioned in article. Why aren't academic articles internally dated?)
NB: That's one paper I've found, not necessarily the most applicable, though it points in the right direction. I'll update (edit or reply) if I find something more appropriate.
Taking threads that trigger the flamewar detector (simple comments vs points comparison) and then running them (either individual comments or the comments as a whole) through an LLM to score them for flamebait seems obvious so my question is why didn't that go anywhere. If you didn't want to use new technology, you could still just build a classifier based on the usernames involved. It wouldn't have to be perfectly accurate, just enough to flag something for manual review.