Learning Concept Descriptions in a Growing Language

Abstract

Humans are capable of producing compact high-level concept descriptions built from previously known concepts and attribute values. In the method presented here, initially concepts are described in terms of attribute values. These descriptions are in a probabilistic DNF form. Assuming a growing language, concepts already known to the system can be used in describing new concepts. The order of teaching the concepts is the key to producing their optimal descriptions. By “optimal” we mean the minimum number of occurrences of constants in descriptions. The problem of finding the minimal description for each concept is NP-complete, hence our proposed algorithm has to be heuristic. Our strategy is based on clustering terms in concept descriptions in order to replace them by shorter higher level terms. Results of the algorithm are optimized concepts descriptions in terms of a growing language, and a concept network that can be used for further learning and reasoning within the concept knowledge base.