Then perhaps use sentiment analysis to remove the derailing part of both negative and negative comments?

i.e.,

"I think these questions at IAC are great,..."

"...they should be asking about <fascinating technical detail>"