New publication at ESWC 2026: Geometric embeddings for EL++ ontologies with deductive-closure-aware training
A new paper by Olga Mashkova, Fernando Zhapa-Camacho and Robert Hoehndorf, presented at ESWC 2026, improves geometric ontology embeddings by training them with the deductive closure and a richer treatment of negative samples.
About
The Bio-Ontology Research Group presented "Enhancing Geometric Ontology Embeddings for EL++ with Negative Sampling and Deductive Closure Filtering" at the 23rd Extended Semantic Web Conference (ESWC 2026). The paper, by Olga Mashkova, Fernando Zhapa-Camacho and Robert Hoehndorf, addresses two long-standing weaknesses of geometric embedding methods for ontologies written in the EL++ description logic.
Geometric ontology embeddings map classes, relations and individuals into a latent space, where similarity can be computed and new axioms inferred. Existing methods based on high-dimensional ball representations have shown promise, but they typically (i) do not distinguish between statements that are unprovable and statements that are provably false (and so may sample entailed axioms as negatives during training), and (ii) make no use of the ontology's deductive closure to identify statements that are inferred but not asserted.
The paper introduces negative-sampling losses that account for the deductive closure and for different categories of negatives, and shows that these modifications consistently improve a family of ball-based ontology embedding methods on the task of knowledge-base completion. The work is part of the BORG group's broader programme on neuro-symbolic methods that bring formal ontology semantics into latent representation learning.
The paper is published in the ESWC 2026 proceedings (Springer LNCS); DOI: 10.1007/978-3-032-25156-5_14.