Inconsistency Monitoring in a Large Scientific Knowledge Base

Abstract

Large scientific knowledge bases (KBs) are bound to contain inconsistencies and under-specified knowledge. Inconsistencies are inherent because the approach to modeling certain phenomena evolves over time, and at any given time, contradictory approaches to modeling a piece of domain knowledge may simultaneously exist in the KB. Underspecification is inherent because a large, complex KB is rarely fully specified, especially when authored by domain experts who are not formally trained in knowledge representation. We describe our approach for inconsistency monitoring in a large biology KB. We use a combination of anti-patterns that are indicative of poor modeling and inconsistencies due to underspecification. We draw the following lessons from this experience: (1) knowledge authoring must include an intermediate step between authoring and run time inference to identify errors and inconsistencies; (2) underspecification can ease knowledge encoding but requires appropriate user control; and (3) since real-life KBs are rarely consistent, a scheme to derive useful conclusions in spite of inconsistencies is essential.