Leveraging Non-lattice Subgraphs to Audit Hierarchical Relations in NCI Thesaurus

Auditing National Cancer Institute (NCI) thesaurus is essential to ensure that it provides accurate terminology for cancer-related clinical care as well as translational and basic research. We leverage a structural-lexical approach to identify missing hierarchical IS-A relations in NCI thesaurus based on non-lattice subgraphs and derived lexical attributes of concepts. For each concept in a non-lattice subgraph, we use two ways to derive the concept's lexical attributes: (1) inheriting lexical attributes from its ancestors within the subgraph; and (2) inheriting lexical attributes from all its ancestors. For a pair of concepts not having a hierarchical relation, if the lexical attributes of one concept is a subset of that of the other, we suggest there is a potential missing IS-A relation between the two concepts. Our approach identified 547 non-lattice subgraphs in the 19.01d release of NCI thesaurus which revealed a total of 1,022 unique potential missing IS-A relations. A random sample of 100 relations was evaluated by a domain expert. Among these relations, 90 can be obtained by the way of inheriting lexical attributes from ancestors within non-lattice subgraph, among which 76 were confirmed as valid (a precision of 84.44%); and 82 can be obtained by the way of inheriting all ancestors, among which 73 were confirmed as valid (a precision of 89.02%). The results show that our structural-lexical approach based on non-lattice subgraphs is effective for auditing NCI thesaurus.