Exploiting regularity in Cyc for text generation

In order to provide hypertext with text generation facilities we combine hypertext with knowledge bases and create expertext systems. This paper studies a particular characteristics of hierarchical semantic nets called regularity, which has proved to be a very valuable tool for expertext construction and use. We show how regularity can be transposed in the context of the Cyc knowledge base to provide an operational tool for understanding the knowledge base intimate topology. Using radical and provoking hypothesis i) general if-then rules express irregularities, and ii) "the more regular, the less interesting", we associate irregularity to text units relevance, to assist expertext users in a number of way, including finding relevant places to input text, helping formulate reading or writing plans, understand the overall system, and suggesting modifications to the knowledge base designers. 1. From hypertext to expertext The combination of hypertext technologies with expert systems techniques is called "expertext system" [Rada 89]. Following [Streitz 89] and [Wand & al 91], we think that writing is a complex problem solving activity whose production can be viewed as an external representation of internal knowledge structures of the author. This study is part of a general study on building expertext from existing knowledge bases, aimed at studying the nature of writing plans [Rada & Barlow 89] as hypertext traversal strategies [Mili&Rada 88, 90a]. We use the Cyc knowledge base [Lenat & Guha 90] as a testbed for experimenting with our techniques, and as a basis for the construction of expertext. This paper is based on two observations resulting from the study of the Cyc system in the context of expertext construction. The Cyc system contains an enormous amount of information and intelligence but uses a variety of knowledge representation techniques which make it hard to understand at a global level. By trying to use the system and manage its complexity, we came to a paradoxical observation concerning the nature of the knowledge embodied in Cyc, i.e., that certain types of knowledge tend to increase regularity, in a precise and computable sense, whereas others tend to break those regularities. Applying regularity calculus to Cyc, the second observation is that for most relations explicitly present in the system, "the more regular" a relation, the less interesting it is as a potential text unit holder. The paper is organized as follows. We will first introduce briefly the Cyc system and its main characteristics with regard to expertext construction, and recall the definition and main results of the regularity calculus as defined in hierarchical semantic nets. We will then apply the notion of regularity to the Cyc system, and show how various inference patterns of the system have corresponding regularity properties. Finally we discuss the application of the regularity hypothesis to expertext construction.