I think content addressing might be a way to go. I.e. address an object (or ontology) by the hash of its content. One could use a special URN namespace (draft-thiemann-hash-urn-01) so that the URI would be something like urn:hash:abcdefffabcdeabf.
The problem is how to normalize data in a sensible way that it can be hashed consistently. InfoCentral is an interesting project working on this and related issues.