EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

semantic-swingers at LLMs4OL 2026 Tasks Flagship, Reuse, and Taxonomy: Retrieval, Fine-Tuning, and the Seen-Unseen Gap in LLM Ontology Learning

Matias Vizcaı́no, Jagdsh Chand, Andy Symonds

Open Conference Proceedings · 02.11.2026 · 11

DOI: 10.52825/ocp.v11i.3681 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Open Conference Proceedings, online erschienen am 02.11.2026. Verfasst von Matias Vizcaı́no und 2 weiteren Autor:innen. Die Arbeit nennt 20 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Ontology learning turns raw text into a structured knowledge graph of terms, types, and their relations. We study why this is hard under exact string match against a curated ontology, on ontologies never seen in training. Our central finding is that success is decided less by reasoning than by surface form: the model usually picks the right concepts and relations, but is scored wrong because the entity name it writes (say, “sweet riesling”) does not match the ontology’s canonical spelling (“SweetRiesling”) – a form that never appears verbatim in the source. We frame this surface-alignment bottleneck as a falsifiable theory – the right vocabulary must be supplied at generation time, through fine-tuned weights or retrieved examples, and was not recovered by any of the after-the-fact repairs we tested – and test it with five hypotheses across the three LLMs4OL 2026 tasks. Every after-the-fact repair we tried (canonicalization, logical closure, self-consistency, keep/drop filtering) failed; among the levers we tested, only retrieval and fine-tuning reliably improved performance on unseen vocabularies. For taxonomy discovery we additionally contribute a scalable structural-matrix alternative that bypasses the LLM, avoiding per-node inference blow-up on very large ontologies. Strikingly, a fine-tuned open 9B model beat a stronger paid API on unseen extraction, even with no retrieval at inference – the bottleneck is which vocabulary the model learned, not raw capability. We contribute a generalization-first evaluation protocol (a vocabulary-disjoint holdout used to reject methods whose gains do not survive on unseen vocabularies) and a catalogue of negative results other teams can skip.

Zitieren

APA

Vizcaı́no, M., Chand, J., & Symonds, A. (2026). semantic-swingers at LLMs4OL 2026 Tasks Flagship, Reuse, and Taxonomy: Retrieval, Fine-Tuning, and the Seen-Unseen Gap in LLM Ontology Learning. Open Conference Proceedings, 11. https://doi.org/10.52825/ocp.v11i.3681

Vancouver

Vizcaı́no M, Chand J, Symonds A. semantic-swingers at LLMs4OL 2026 Tasks Flagship, Reuse, and Taxonomy: Retrieval, Fine-Tuning, and the Seen-Unseen Gap in LLM Ontology Learning. Open Conference Proceedings. 2026;11. doi:10.52825/ocp.v11i.3681

BibTeX
@article{Vizcano2026,
  author = {Vizcaı́no, Matias and Chand, Jagdsh and Symonds, Andy},
  title = {semantic-swingers at LLMs4OL 2026 Tasks Flagship, Reuse, and Taxonomy: Retrieval, Fine-Tuning, and the Seen-Unseen Gap in LLM Ontology Learning},
  journal = {Open Conference Proceedings},
  year = {2026},
  volume = {11},
  doi = {10.52825/ocp.v11i.3681}
}
Kostenlos starten →