Refining concepts by machine learning

Abstract

In this paper we deal with machine learning methods and algorithms applied in learning simple concepts by their refining or explication. The method of refining a simple concept of an object O consists in discovering a molecular concept that defines the same or a very similar object to the object O. Typically, such a molecular concept is a professional definition of the object, for instance a biological definition according to taxonomy, or legal definition of roles, acts, etc. Our background theory is Transparent Intensional Logic (TIL). In TIL concepts are explicated as abstract procedures encoded by natural language terms. These procedures are defined as six kinds of TIL constructions. First, we briefly introduce the method of learning with a supervisor that is applied in our case. Then we describe the algorithm 'Framework' together with heuristic methods applied by it. The heuristics is based on a plausible supply of positive and negative (near-miss) examples by which learner's hypotheses are refined and adjusted. Given a positive example, the learner refines the hypothesis learnt so far, while a near-miss example triggers specialization. Our heuristic methods deal with the way refinement is applied, which includes also its special cases generalization and specialization.