I am not sure that this will work anytime soon, even if NLP gets better substantially. I think the problem reduces to examination of patents and their novelty. I think that having algorithms to find related work is going to be hard to the point of it being unreliable. For example, in CS there is a trend of changing nomenclature every n years, which I think makes it hard to find related work.
Google manages to serve advertisements despite changing nomenclature. We're just talking about "advertising" papers based on various data about the paper under submission.
