Generating Natural Language Definitions from Classification Hierarchies

Abstract

In interactions with users, knowledge based systems are often called upon to define their terms or concepts [Maybury, 1989]. These terms and concepts usually comprise classes within some classification scheme (e.g., generalization hierarchy). Beyond simply retrieving the superclass of the to-be-defined class (e.g., a mammal is vertebrate) more sophisticated definition also requires selection of distinguishing features or characteristics of this class (e.g., a mammal is vertebrate that gives live birth to and nurses its offspring). To do this, we have refined and extended set theoretic, feature-based models of object similarity and proWtypica1ity, and developed an algorithm that selects the most distinguishing set of attributes and attribute-value pairs of class in the context of taxonomy of classes and their properties based on notions of prototypicality and discriminatory power. In this paper, we illustrate classificatory representation using objects and attribute-value pairs in test domain of vertebrates; describe our algorithm for computing prototypicality, discriminatory power, and distinctive power, based on this sample representation; and show how this algorithm is implemented to generate definitions of object classes in this representation.