Natural language directed inference from ontologies

This paper presents an investigation into the problem of content determination in natural language generation (NLG), using as an example the problem of determining what to say when asked “What is an A?”, where A is a concept defined in an OWL ontology. It is shown that a naive approach to this problem, which just presents a set of the stated axioms, will often inadvertantly violate maxims of cooperative conversation. What is required instead is a kind of inference that generates logical conclusions of the axioms that are suitable for natural language presentation—natural language directed inference (NLDI). Although NLDI, in this case a kind of non-standard inference in description logics, is hard to formalise in general, for this problem we isolate a significant subproblem—that of enumerating subsumers of A that are suitable for natural language presentation. For this problem, which on the face of it appears intractable, we show how factors relevant to natural language presentation enable an optimised solution that is realistic in practice.The paper makes a contribution to the increasingly important practical problem of explaining concepts in an ontology. It also makes a first step towards the development of domain independent principles for content determination.