With knowledge graphs increasing in popularity, aligning and integrating them is paramount to ensure their usefulness and reusability. A key step in this process is ontology matching, whereby the semantic models of KGs are aligned into a single cohesive semantic backbone. While finding simple pairwise equivalences between entities in two ontologies is well addressed by state-of-the-art algorithms, finding more complex mappings that can include multiple entities from different ontologies is far from solved, despite their importance in ensuring a deep and meaningful integration of KGs.

We propose a novel complex ontology matching approach that explores geometric operations over the shared semantic space afforded by large language models, enabling the discovery of complex mappings that are missed by purely lexical approaches. We evaluate our approach on several biomedical ontologies using partial reference alignments and manual expert validation. Our approach improves on the performance of a purely lexical approach while also increasing the coverage of complex multi-ontology alignments by 20 to 80%, which translates to a 97% coverage of the source ontologies. Moreover, the manual evaluation of the mappings produced by LLM shows that it achieves a high level of precision. This work demonstrates that the use of LLMs can improve on the performance of traditional lexical strategies.