Semantic refinement and error correction in large terminological knowledge bases

摘要

Capturing the semantics of concepts in a terminology has been an important problem in AI. A two-level approach has been proposed where concepts are classified into high-level semantic types, with these types constituting a portion of the concepts’ semantics. We present an algorithmic methodology for refining such two-level terminologic networks. A new network is produced consisting of “pure” semantic types and intersection types. Concepts are uniquely re-assigned to these new types. Overall, these types form a better conceptual abstraction, with each exhibiting uniform semantics. Using them, it becomes easier to detect classification errors. The methodology is applied to the UMLS.