From Machine Learning to Deep Learning in Anatomy Ontology Matching: The VIEWAnatomy-ML and VIEWAnatomy-DL Systems at OAEI 2026

This paper presents VIEWAnatomy, an interactive ontology alignment framework evaluated in the Anatomy track of OAEI 2026, in two variants. VIEWAnatomy-ML uses lexical similarities, K-Means and an ensemble of Random Forest, Naive Bayes and Multilayer Perceptron; it reaches precision 0.9640, recall 0.6365 and F-measure 0.7668 (1,001 correspondences). VIEWAnatomy-DL extends this pipeline with Sentence-BERT embeddings loaded via the Deep Java Library (DJL) and a global bidirectional 1:1 filter; it reaches precision 0.9509, recall 0.7672 and F-measure 0.8492 (1,223 correspondences). The ML→DL transition thus yields +0.1307 in recall and +0.0824 in F-measure at a cost of only −0.0131 in precision. Among the 14 official systems plus the two variants (16 entries), VIEWAnatomy-ML ranks 4th in precision and VIEWAnatomy-DL ranks 8th in precision and 11th in F-measure. The comparison shows that generic SBERT embeddings are an effective low-cost step from classical ML to deep learning , while domain-specific knowledge (SapBERT, UMLS) and LLM oracles remain the key differentiators of top systems. A roadmap targeting F-measure in the 0.955-0.960 range for OAEI 2027 is proposed.