"Knowledge, in the form of new concepts and relations in the semantic network, must be supplied by human annotators.. We can refine the learn() function into an unsupervised, bootstrapped learning mechanism."

I'm really interested in that second option of unsupervised concept-relation-graph learning, are there any good pointers to prior art?

I think the problem naturally fits probabilistic graphical models but existing PGM algorithms are way too complex to scale to real data. On the other hand deep learning rarely goes beyond object recognition.