Understanding recipes semantically requires identifying food entities and linking them to ontology concepts. Food mentions in recipes are often ambiguous, context‑dependent, and lexically diverse, making ontology entity linking (EL) difficult. SCODEL introduces a model‑agnostic framework whose central idea is that contextual food mentions and ontology concepts referring to the same entity should learn an identical canonical ontology definition. These shared definitions act as a semantic normalisation layer, supplying complementary information to existing EL models without altering their architecture. We also release ReciLinkD, the first benchmark for contextual food ontology EL in recipes. Experiments show that SCODEL consistently boosts three modular EL models: on seen concepts SCODEL × BLINK achieves the highest Hit@1 of 88.33%, while on unseen concepts SCODEL × GSM‑EL reaches the highest Hit@1 of 83.15%, both surpassing the strongest large language model (LLM) based EL baseline. Additional tests on a biomedical ontology benchmark confirm SCODEL’s generalisability beyond the food domain. Finally, we apply SCODEL at scale to ground food mentions and build an ontology‑grounded recipe knowledge graph, demonstrating its usefulness for ontology‑aware retrieval and corpus‑level food analysis.
Review: SCODEL’s use of shared canonical definitions provides a clean semantic bridge that enhances diverse EL systems without architectural changes, especially excelling on unseen concepts and showing promising cross‑domain transferability.