Over the past decades, the vision of the Semantic Web has captivated researchers: a web where information is not just readable by humans but also interpretable by machines. Central to this vision are ontologies—formal, structured representations of knowledge that define concepts, relationships, and rules within a domain. They allow systems to reason, integrate knowledge across sources, and reuse information effectively.

However, building these ontologies has historically been slow, manual, and expert-driven. Domain experts must carefully identify terminology, define hierarchies, and model semantic relationships. As scientific and technological knowledge grows—particularly in biomedical research, materials science, and scholarly communications—the gap between the rapidly growing body of unstructured knowledge and the structured, machine-readable ontologies has broadened dramatically.

Advancements in Large Language Models (LLMs) present a significant opportunity to accelerate ontology learning. LLMs can process massive amounts of text, suggest terms, classify entities, and even propose relationships—tasks that traditionally required months of expert effort. Yet, automation alone is not enough. To trust and scale LLM-assisted ontology construction, the community needs:

• Systematic evaluation of generated ontologies.

• Reproducible infrastructures to benchmark and compare approaches 

• Shared datasets and tasks to push research forward 

This is where the LLMs4OL Challenge comes in!

Watch our short promotional video below to learn more about our mission.

From Research to Community Infrastructure

Three years of growing a shared ecosystem for ontology learning with large language models — from a foundational LLMs4OL tasks paradigm to a challenge series, now to an open toolkit.

In 2023, our paper "LLMs4OL: Large Language Modes for Ontology Learning" (ISWC 2023) demonstrated that LLMs hold remarkable potential to automate the most labor-intensive OL subtasks — from typing lexical terms to discovering taxonomic hierarchies and extracting non-taxonomic relations. What started as an empirical study became a challenge series: first at ISWC 2024 (Baltimore, USA), then at ISWC 2025 (Nara, Japan), attracting a growing community of researchers across NLP, Semantic Web, and ML. Now, the 3rd edition introduces OntoLearner — a modular Python toolkit that finally gives the community a unified, reproducible infrastructure for OL experimentation at scale.  

This year, in the 3rd LLMs4OL Challenge @ ISWC 2026, participants will tackle three visionary scenarios: 1) building ontologies end-to-end from raw text, 2) extending existing ontologies, and 3) learning taxonomies that generalize across domains. Together, these tasks reflect real-world demands where knowledge is dynamic, interconnected, and domain-spanning. We invite researchers to push the limits of LLMs in constructing, evolving, and transferring structured knowledge. 

💡💡  Visit the challenge overview page for more information on the 3 primary tasks of the challenge, and also an optional task.