When Context Is Not Enough: Structural, Compositional, and Contextual Interventions in Biomedical Terminology Processing

While biomedical terminology systems must identify term spans, link mentions to concepts, and generate definitions, retaining additional structural alternatives or contextual evidence may not improve decisions around these elements. We evaluated three interventions on audited specialized data with locked internal holdouts. A PubMedBERT neural boundary model was paired with nested or non-overlapping term decoding across five seeds; a rule-based concept linker compared open normalization with lexical decomposition for concept mapping. Additionally, for definition generation, GPT-4o-mini received either a concept-only prompt or occurrence passages selected in stable order or ranked using training-derived PubMedBERT embeddings and K-means clusters. Non-overlapping decoding produced a large, consistent increase in term-occurrence micro-F1 across all five neural seeds. Nested decoding recovered more annotated occurrences but produced many false-positive multiword spans. On the concept-linking test, open normalization achieved a micro-F1 of 0.8110 versus 0.7829 with decomposition, although decomposition helped multi-label cases. On the definition test, concept-only generation had the highest descriptive combined score as compared with prompts including unranked and ranked evidence. The cluster-derived score changed decisions either in the opposite direction or without reliable gains. These benchmark-relative findings indicate that richer representations require selection rules aligned with the target decision at play.