OntoLearner: A Modular Python Library for Ontology Learning with Large Language Models

Ontology learning (OL) aims to bridge the gap between unstructured text and structured, machine-interpretable knowledge models. Despite decades of research, the field has remained fragmented, lacking a unified infrastructure for systematic evaluation. Existing tools often operate in isolation, focusing either on ontology hosting or specific extraction methods, without a shared framework to measure progress across domains or model architectures.

OntoLearner is a modular, cross-domain Python library designed to unify ontology access, LLM-driven learning pipelines, and standardized benchmarking. It provides four primary contributions: a unified interface for transforming diverse ontologies into machine-readable formats; a modular library supporting core OL tasks (term typing, taxonomy discovery, and non-taxonomic relation extraction) with standardized datasets; integration of LLMs as assistive components; and an extensible platform aligned with FAIR principles. By hosting 180 ontologies across 22 domains on HuggingFace, the framework enables reproducible, large-scale empirical studies of OL methods.

Using this infrastructure, the authors conducted a large-scale evaluation of 22 retrieval models and 12 LLMs. The results demonstrate that performance is not simply a function of model scale or architectural sophistication. Instead, the authors identify a fundamental structural mismatch: models often struggle to encode hierarchical entailment (is-a relations) because their embedding spaces collapse under dense, compositional labels, and they suffer from hubness effects. Furthermore, error patterns—such as systematic overprediction or underprediction—are determined by the intrinsic complexity of the ontology (e.g., class density and relational depth) rather than the model's parameters. This suggests that the bottleneck in OL is the structural incompatibility between neural knowledge encoding and symbolic ontology organization.