Learning Implicit Editorial Rules for Medicinal Terminology Naming Using Rule-Compositional Cascading Prompts

Medicinal terminologies depend on expert-authored fully specified names (FSNs) and preferred terms (PTs) that encode implicit editorial rules not formally represented in ontologies. This work explores the use of large language models to learn and apply these rules for automated medicinal terminology naming. Using existing OWL-defined concepts from the HSE Ireland medicinal terminology, this project evaluated a four-level approach ranging from direct fine-tuning to rule-compositional cascading prompts. Results across six medicinal concept types show that the cascading prompt framework further improves accuracy, numerical robustness, and interpretability by explicitly modelling editorial decision steps, enabling scalable and consistent FSN and PT generation.