What Stefan describes on this old post still holds true - we’re not currently able to distinguish potentially useful keywords from irrelevant ones.
I’m not sure how feasible this would be to implement, but the matter continues to be on our radar so I’ve added your +1 to the topic.