Building and maintaining hierarchical semantic nets
Hafedh Mili, Simon Berkovitch · 1988
Knowledge intensive AI applications require the development and maintenance of increasingly large, consistent and up-to-date knowledge bases. Knowledge engineering remains largely manual, presenting a major bottleneck to the development of intelligent systems. This research explores methods to automatically build and maintain hierarchical semantic nets. Such semantic nets exist that are manually built and laboriously updated according to some rules that are imprecise at best. The approach followed is one of taking advantage of the structure existing in readily available knowledge sources to build and/or maintain hierarchical semantic nets. Three methods were investigated, dealing with knowledge sources of varying structural complexity. The first two methods, briefly summarized in this thesis, proved useful in the context of an intelligent information retrieval system that used a hierarchical semantic net to match documents to queries. However, both methods suffered, to varying degrees, from the vagueness and lack of clear semantics of hierarchical relations in man-made hierarchies. The third method, DK method, is based on a precise characterization of hierarchical relationships. The DK method is based on the observation that man-made hierarchies are often built in a way such that concepts' properties follow clear patterns, similar to the patterns implied by inheritance in taxonomies. These patterns are described by regularity, a generalized form of inheritance. It is argued that regularity is a fundamental property of hierarchies, and a model of hierarchies (DK model of hierarchies) that embodies regularity is proposed. In this model, hierarchical relations are described by predicates that represent the particular relationships that hold between the properties of a concept and the properties of its descendants. Adding a concept to a DK hierarchy is then reduced to a problem of classifying that concept using the hierarchy as a hierarchical classifier. The cognitive and mathematical foundations of this approach are discussed. In particular, a fuzzy model of DK hierarchies and its associated classification algorithm are proposed to handle the uncertainties and exceptions to regularity that often plague man-made hierarchies. The fuzzy models are then tested on two man-made hierarchies, and the results show the validity of our approach.