Well, automated systems constructed by people - as ontology construction and semantic reasoner writing is primarily a human activity (... yeah I see it). It would be pretty funny if we end up with a semantic web, where logic is king, having ontologies mapped by a probabilistic AI though.
As far as edge cases are concerned, that is why you use an open world assumption [0], allowing the reasoner to return "Unknown". Which would act as a compiler error for human understating. Of course that won't save you from reclassifications, but nothing will. I haven't done much thinking on the matter, but I can't think of anything that would better handle the expansion of a knowledge graph than an OWA semantic reasoner. I've done a bit of work with neural networks, and changes in the domain are not well handled by neural networks. I imagine the results of a large graph change would create in the neural network something that would look very human - confirmation bias :)
If you've set things up with that sort of constraint, then you've set yourself up for failure - Gödel's incompleteness theorems [0]. I think OWA is the only way around that, you just need to gracefully handle unknowns. Of course, you can convert an unknown into a known pretty easily with: ¬u :)
> For instance in the case of cardinality problem.
That is exactly when I'd want my reasoner to ignore a graph edge, or else you get a non sequitur that ends with Nike belonging in the human set. Measure twice, cut once.
[0] http://en.wikipedia.org/wiki/G%C3%B6del%27s_incompleteness_t...
F=m*a is a great rule most of the time that happens to stop working at various edge cases. Sadly, you can say the same of all of physics and just about every other rule. Even things as 'simple' as living vs dead can get really complicated with lots of odd edge cases.
http://en.m.wikipedia.org/wiki/HeLa new species?