Revision of logical theories

This paper presents a system for revising hierarchical first-order logical theories, called INCR/H. It incorporates two refinement operators, one for generalizing clauses which do not cover positive examples, and the other one for specializing the inconsistent hypotheses inductively generated by any system that learns logical theories from positive and negative examples. The generalizing operator is inspired from Rayes-Roth and McDermott's Interference matching, while the specializing operator is completely novel. Both of them perform a search in the space of logical clauses and take advantage of the structure of this set. The main characteristic of the system consists of the capability of autonomously performing a representation change, that allows INCR/H to extend the search to the space of clauses with negated literals in the body (program clauses) when no correct theories exist in the space of definite clauses. Experimental results in the area of electronic document classification show that INCR/H is able to cope effectively and efficiently with this real-world learning task.