This paper presents the results achieved by BioSTransMatch in the OAEI 2025 competition, marking its second participation in the Bio-ML track. Unlike the 2024 version, which combined supervised and unsupervised learning, the 2025 edition focuses exclusively on a graph-based unsupervised approach. The proposed method exploits the hierarchical structure of SNOMED CT by generating graph-based contexts through parent-child relations. These contexts are then used to train sentence transformer models using positive and negative pairs derived from the ontology graph. We trained three diferent transformer backbones and evaluated their performance on the equivalence and local ranking subtasks of the Bio-ML track. The obtained results highlight the potential of structural context learning for ontology alignment without explicit supervision.
BioSTransMatch aims to explore representation learning directly from biomedical ontology graphs. The goal is to create concept embeddings that encode not only lexical semantics but also hierarchical relationships such as is-a links, without using manually aligned pairs. This allows the matcher to adapt to new ontologies with minimal preprocessing.
To introduce structural knowledge, we transformed the SNOMED CT ontology into a directed graph, where nodes represent concepts and edges represent hierarchical relations. For each node, all direct parents and children were collected to build a contextual description as follows:
parent of <parent_label> ; child of <child_label>
Each concept–context pair is treated as a positive example, while a random concept from the graph serves as a negative pair. The model is trained using a Cosine Similarity Loss, which maximizes similarity between positive pairs and minimizes it between negatives. Formally, for embeddings u and v, the similarity is computed as: (, ) =
.
|||||||
Training follows the standard SentenceTransformer pipeline with a batch size of 16, one training
epoch, and intermediate evaluation using an embedding similarity evaluator. We used our model
BioSTransformers [
After encoding ontology concepts into embeddings, we construct a similarity matrix based on cosine similarity. Each row corresponds to a source concept and each column to a target concept. For the global matching task, the highest-scoring target is selected for each source. A similarity threshold of 0.75 is applied to remove low-confidence correspondences (See Figure 1).
The Bio-ML track1 consists of five diferent pairs of datasets and includes both an equivalence matching task and a subsumption matching task. BioSTransMatch participates in the equivalence matching task only.
The ontologies of this track are the OMIM (Online Mendelian Inheritance in Man), ORDO (Orphanet Rare Disease Ontology), NCIT (National Cancer Institute Thesaurus), DOID (Human Disease Ontology), 1https://www.cs.ox.ac.uk/isg/projects/ConCur/oaei/ FMA (Foundational Model of Anatomy), and SNOMED CT (Systematized Nomenclature of Medecine Clinical Terms).
• OMIM desribes genes, genetic phenotypes, and gene-phenotype relations, generated through
manual curation based on biomedical literature [
The equivalence matching task is further divided into two categories: an unsupervised setting, and a semi-supervised setting, where 30% of the reference alignments are provided in the training set.
In this experiment, the model was applied according to the workflow illustrated in Figure 1. Table
1 presents the results on the SNOMED–FMA task, both with and without context. The context-free
(without context) configuration corresponds to the first participation of BioSTransMatch [
Local Ranking In this step, we take all candidates and sort them in a decreasing order.
Training the model with contextual information led to a slight improvement in the equivalence matching task and a notable improvement in the ranking task. This indicates that incorporating contextual information helps the model capture a richer semantic understanding between entities. While the model performs well in local ranking, it still struggles to identify exact equivalences. These results highlight the need for further work to enhance its ability to discriminate true matches among similar candidates, possibly by exploring alternative strategies for selecting the best candidate in equivalence matching rather than relying solely on the maximum score.
The 2025 edition of BioSTransMatch introduces a fully graph-driven unsupervised approach for ontology matching. By training on contexts generated from hierarchical walks within SNOMED CT, the system learns representations that capture both textual and relational semantics. Despite the absence of supervision, the results are competitive, particularly in the ranking task.
Future work will focus on integrating multi-hop relational contexts, combining textual definitions with structural paths, and exploring contrastive graph objectives to enhance structural embeddings. This approach opens promising directions for scalable, domain-adaptive ontology matching.
The author(s) have not employed any Generative AI tools. the 19th International Workshop on Ontology Matching co-located with the 23rd International Semantic Web Conference (ISWC 2024), Baltimore, USA, volume 3897, 2024, pp. 132–137.