Machine learning on that corpus would be extremely difficult task. This is what engineers don't really get about patent applications. We see the underlying technology or specifications and say, same thing-identical.

IP Lawyers, patent clerks look at the claims and see difference. Claims have two functions. Public notice and kind of a "patent scope" of applicability. It's all very obtuse and would not be very easy I think to attach semantics via ML alone.

There have been some crowdsource attempts: http://peertopatent.org https://www.articleonepartners.com