Iterative Re nement of Knowledge Bases with Consistency Guarantees

Natural kinds, such as the concepts toy block or photosynthesis, are ubiquitous in human reasoning yet lack deenitional (i.e., individually necessary and jointly suucient) properties as membership conditions. We want to represent a wide variety of assertions about these natural concepts with precision and logical consistency. Furthermore , we want to be able to verify that the meaning of each concept does not connict with the constraints implied by its constituents. Classiication-based approaches to these tasks have relied on deenitional properties in order to deduce subsumption and thus detect inconsistency. These approaches are inapplicable for building representations of natural concepts. Our solution to these problems is twofold. First, we build KBs by iterative reenement; i.e., the KB is constructed through a sequence of editing operations. Second, we deene practical, yet formal , properties of concept satissability and inheritance that each operator is guaranteed to preserve. When the user builds representations of concepts using the operators, guard procedures ensure that existing concepts are used in a manner consistent with their meaning. We also discuss the computational complexity of guarding consistency, and the characteristics of the system that we believe make these computations practical in an interactive setting.