Chapter 8 Regularity , Document Generation , and Cyc

We are interested in building and maintaining semantic nets in general, and hierarchical semantic nets in particular. In [Mili, 1988], we developed a model of hierarchical semantic nets that generalizes taxonomic models by replacing the concept of property inheritance by a more general behavior of properties that we called regularity [Mili & Rada, 1990a]. We are also interested in the generation of structured documents from hypertext. We argue that authors structure their descriptive documents based on personal but systematic traversals of a model of the domain of discourse [Mili & Rada, 1990b]; when that model consists of a semantic net that exhibits regularity, the traversal can be rationalized and described concisely. Within the context of our research, we acquired an electronic copy of the Cyc knowledge base [Lenat & Guha, 1990] to: 1) use the semantic net underlying Cyc to support the generation of argumentative and explanatory documents, and 2) identify regularity patterns. Interestingly, the sheer size and richness of Cyc posed challenging performance problems to its designers, who had to constrain proof and inference procedures in such a way as to make it ill-adapted to 1LARC, in French, stands for Laboratoire pour l’Acquisition et la Représentation des Connaissances. the kind of open-ended lengthy logical inferences required for generation of argumentative text. The study of regularity patterns in Cyc led us to generalize the concept of regularity, and to formulate a number of hypotheses about the semantic structure of Cyc, and of common sense knowledge in general, proving once more that regularity is a powerful tool for managing the complexity of large knowledge bases.