Learning Constrained Atoms

This paper studies the generalization of atomic formulas, or atoms, that are augmented with constraints on or among their terms. The atoms may also be viewed as definite clauses whose antecedents express the constraints. Atoms are generalized relative to a body of background information about the constraints. The paper develops an algorithm for the generalization task and discusses algorithm complexity. The paper also presents semantic properties of the generalizations computed by the algorithm. We have shown elsewhere that these properties are useful in problems such as abduction, induction, analogical reasoning, and knowledge base vivification. This paper emphasizes the application to induction and presents a paclearning result for constrained atoms.

Duke Scholars

Author David Page Biostatistics & Bioinformatics, Division of Biostatistics