Towards an Abstraction-Refinement Framework for Reasoning with Large Theories