We are engineering agent communication that tolerates semantically heterogenous terminology. Unlike methods for integrating databases, we do not translate to a central, shared language. Instead, agents communicate directly, but imperfectly. We restrict semantic heterogeneity to a form conducive to decentralized development of open societies of agents. In differentiated ontologies, concepts are not shared, but inherit structure from concepts that are shared. Rough mapping identifies syntactic similarity and differences between expressions in differentiated ontologies. This algorithm works by maximizing structural isomorphism between graphs of the expressions.
We will test our algorithms with simulations. These will generate artificial ontologies with denotations in a set of possible worlds. Syntactic mapping evaluation measures structural isomorphism. Semantic evaluation, in our finite simulated worlds, calculates overlap of the mapped concept extensions. Strong correlation between syntactic and semantic evaluations will demonstrate the validity of our mapping algorithms. Furthermore, we will quantify the relationship between the degree to which agents share language, and the probability of effective communication. In a preliminary exercise, high-quality mappings were built between expressions that share only a third of their definitional structure. The simulations will also support development of methods for translating agent requests (rather than mapping them to existing expressions), and for inducing whether specific mappings will be successful in particular problem contexts.
This research will contribute guidelines for language development in open societies of agents; automated mapping between ontologies using new algorithms for analogical reasoning; and a method for experimenting with computational communication under conditions of semantic heterogeneity.
My dissertation proposal (1997). Available by request.