Most existing work on concept formation (e.g., conceptual clustering) has focused on the formation of concepts for classification purposes, resulting in an emphasis on producing sets of concepts that are complete, i.e., cover all observed instances. In this research, we are primarily concerned with concept formation as a solution to the new term problem, which results in a new set of requirements on concept formation: concept formation must be triggered in a demand-driven fashion by the shortcomings of the existing representation, and a concept's quality must be evaluated in terms of its contribution to the representation. The approach to demand-driven concept formation that we present in this paper is part of an integrated learning system called the MODELER, the learning component of the knowledge acquisition system BLIP, and is capable of forming concepts incorporating both disjunctions and relational information. Much of the power of the approach stems from its close integration with the other components of the learning system, namely, rule discovery and knowledge revision, which are also presented in this paper.