I think that given the complexity of the problem they don't even try to parse the sentence and do attribution - they just shotgun with ML and hope for the best.