IMHO, a more practical flow would be: warn the user 3 times, but on the 3rd & final time state that : this will be submitted for review by a human reviewer & if found abusive, your a/c will be disable for x days.
A genuine commenter would gladly agree to such a request.
FB has humans who could review this, or they could use Amazon's Mturk - ship off each abusive comment as a HIT which gets posted after the 3 warning attempts, & maybe send it to multiple reviewers(mturk workers) - & use a majority vote, to decide among the reviews.
Decision could be binary ( ban/no ban) or some reputation metric ( decrement by 10 points , & if no suspect activity occurs in the next x(30?) days, karma/ reputation to be restored to the original number.