Auto Ontology: Towards Automated Term-to-Concept Assignment in Microbiology Analytics

Maintaining clinical ontologies requires continuous addition of new synonyms from laboratory information systems, which is currently a manual and time-consuming process. This study evaluates whether text embedding models can support semi-automated synonym-to-concept mapping for ontology maintenance. Using the Momo microbiology ontology, we indexed term embeddings locally with Qdrant and compared semantic similarity for assigning unseen terms to known concepts. Performance was assessed using Top-k recall, Mean Reciprocal Rank, and F1-score. Results show that dense embeddings, particularly BioLORD-2023-M (Top-1 recall = 0.68, F1 = 0.67), can retrieve relevant concept suggestions with sufficient accuracy for human-validated workflows, albeit not yet for automatic assignment. This approach demonstrates the potential of embedding-based semantic retrieval for supporting practical ontology curation in resource-constrained clinical settings.