A new pipeline uses GPT-4o to spot valid connections between scientific ontologies that similarity math alone keeps missing.
Researchers built an end-to-end system to automatically link concepts across 33 ontologies covering machine learning, microscopy, computational science, and experimental workflows, a project called ReproduceMeON. The pipeline first used domain-adapted DistilBERT embeddings and clustering to narrow roughly 800,000 raw concept pairs down to 95,000 plausible candidates. GPT-4o then reasoned about each candidate through iterative prompting to decide which relationships actually made sense and label them with a specific type. Two human annotators checked 429 of the generated links and found 80.19% were correct overall, rising to 91.49% on the system's highest-confidence calls, for an F1 score of 0.890.
That gap matters because the team also tested five similarity-based baselines, including Sentence-BERT, and the best of them scored an F1 of just 0.581, with an ablation showing these methods are barely better than a coin flip (AUC around 0.5) at telling real relationships from noise on the same filtered candidates. Ontologies are the shared vocabularies that let databases, lab instruments, and software agree on what words mean, and stitching them together by hand is a chore mostly well-funded knowledge-graph teams can afford.
If that holds up outside this one test network, it is one more case this year of the same pattern: vector similarity gets you in the neighborhood, but something still has to read the words to tell a neighbor from a match.