Why don't you start by describing how your theoretical UI generation would work?
Then people use the universal query interface to pick a node (abstract or concrete) from the graph.
Then they repeat the step above to pick another node from the graph.
Once they have 2 nodes, they can pick an edge (relationship) that works between those two.
There is your basic interaction, where people can communicate semantic statements.
You can easily expand from there.