Nope - It's the other way around - the schema was described in natural language based on the documents. In this way, you can make sure the schema described and used is based on the documents you are processing, and that you can create a graph based on your own opinion