Computing OWL Ontology Decompositions Using Resolution Robert Schiaffino 3 , Achille Fokoue 1 , Aditya Kalyanpur 1 , Aaron Kershenbaum 1 , Li Ma 2 , Chintan Patel 4 , Edith Schonberg 1 , and Kavitha Srinivas 1 1 IBM Watson Research Center,P.O.Box 704, Yorktown Heights, NY 10598, USA achille, aaronk, ediths, ksrinivs@us.ibm.com 2 IBM China Research Lab, Beijing 100094, China malli@ cn.ibm.com 3 Iona College, New Rochelle, NY 10000 rschiaffino@iona.edu 4 Columbia University Medical Center chintan.patel@dbmi.columbia.edu Abstract. Reasoning over large ontologies can be done more effectively if they can be decomposed into smaller parts which can be reasoned on independently. This requires identifying parts of the ontology relevant to the problem at hand. We present a novel algorithm, based on resolution calculus, for decomposing an OWL ontology into smaller, more manage- able components, such that the union of reasoning over each of these components separately is the same as reasoning over the original ontol- ogy. We describe our computational experience using the algorithm, and demonstrate that it is indeed possible to efficiently solve the standard concept subsumption reasoning problem in four large real-world OWL ontologies: SNOMED-CT, NCI, SWEET-JPL and GALEN. We chose SNOMED-CT and NCI because of their size; Galen because it is highly interconnected; SWEET-JPL because it is expressive (containing nega- tion, both existential and universal quantifiers and both intersections and unions). 1 Introduction Ontologies that model domains in the real world are often very large, pushing the limits of existing reasoners. For example, SNOMED CT 5 , has over 300,000 concepts and approximately the same number of axioms. To reduce the cost of reasoning over large ontologies, we propose a sound and complete technique which allows us to work only with axioms and concepts relevant to the current task. In some cases, this may also involve merging results obtained from parts of a decomposed ontology. In this paper, we focus on OWL ontologies, specifically the TBox which con- tains terminological assertions about concepts and the RBox which contains as- sertions about roles and role hierarchies. Furthermore, we focus on the standard 5 http://www.snomed.org/snomedct/index.html