I wonder if this could be formalized in terms of Bayesian belief networks. Being able to split nodes, or delete deprecated ones, or add entirely new ones, in order to represent improved ontologies would seem like it'd have extremely valuable applications.