Computer-assisted checking of conceptual relationships in a large thesaurus

We describe a method to support quality control of relationship instances in a large thesaurus or other KOS, using the example of AGROVOC (~33K concepts and ~97K conceptual relationship instances), where manually checking each relationship instance is not feasible. Our method identifies relationship instances that should be checked manually; it can also shed light on problems with the definition of relationship types. We apply a simplified version of the linguistic concept of verb valency to the analysis of conceptual relationships, treating relationship types as verbs. We map each of the two concepts in a relationship instance to an entity type; the resulting entity type pair is a valency pattern, as in the following example: Flavivirus < causes> yellow fever→ Valency pattern [microorganism, diseaseOrDisorder A relationship instance that use a valency pattern that is rare for the relationship type might be erroneous and should be checked by an editor. We describe our method in detail, how we associated concepts with the appropriate entity type (this information is not available for AGROVOC) and how we organized the data for analysis. Then we present some illustrative results.