Description Logic Concept Learning using Large Language Models

Adrita Barua, Pascal Hitzler
Proceedings of The 19th International Conference on Neurosymbolic Learning and Reasoning, PMLR 284:160-178, 2025.

Abstract

Recent advances in Large Language Models (LLMs) have drawn interest in their capacity for logical reasoning, an area traditionally dominated by symbolic systems that rely on complete, manually curated knowledge bases represented in formal languages. This paper introduces a framework that leverages pretrained LLMs to generate Description Logic (DL) class expressions from instance-level examples and background knowledge, translated to natural language. The baseline is Concept Induction, a symbolic learning approach that is mostly based on formal logical reasoning over a DL theory. Drawing inspiration from the DL-Learner architecture, our approach replaces traditional symbolic methods with LLM-based models to generate DL class expressions from instance-level data. We evaluate our approach using three benchmark ontologies across two LLMs: gpt-4o and o3-mini. We use a symbolic reasoner, Pellet, to verify the LLM-generated results and incorporate the reasoner’s feedback into our pipeline to ensure logical consistency, thereby generating a hybrid neurosymbolic system. By introducing controlled variations to the background knowledge, we assess the models’ reliance on commonsense versus formal reasoning. Results show that o3-mini achieves near-perfect accuracy across settings, albeit with longer runtime. These findings demonstrate that LLMs have the potential to serve as scalable and flexible DL learners when coupled in a hybrid neurosymbolic setting, offering a promising alternative to symbolic approaches—particularly in contexts where high-quality ontologies are incomplete or unavailable.

Cite this Paper

BibTeX
@InProceedings{pmlr-v284-barua25a, title = {Description Logic Concept Learning using Large Language Models}, author = {Barua, Adrita and Hitzler, Pascal}, booktitle = {Proceedings of The 19th International Conference on Neurosymbolic Learning and Reasoning}, pages = {160--178}, year = {2025}, editor = {H. Gilpin, Leilani and Giunchiglia, Eleonora and Hitzler, Pascal and van Krieken, Emile}, volume = {284}, series = {Proceedings of Machine Learning Research}, month = {08--10 Sep}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v284/main/assets/barua25a/barua25a.pdf}, url = {https://proceedings.mlr.press/v284/barua25a.html}, abstract = {Recent advances in Large Language Models (LLMs) have drawn interest in their capacity for logical reasoning, an area traditionally dominated by symbolic systems that rely on complete, manually curated knowledge bases represented in formal languages. This paper introduces a framework that leverages pretrained LLMs to generate Description Logic (DL) class expressions from instance-level examples and background knowledge, translated to natural language. The baseline is Concept Induction, a symbolic learning approach that is mostly based on formal logical reasoning over a DL theory. Drawing inspiration from the DL-Learner architecture, our approach replaces traditional symbolic methods with LLM-based models to generate DL class expressions from instance-level data. We evaluate our approach using three benchmark ontologies across two LLMs: gpt-4o and o3-mini. We use a symbolic reasoner, Pellet, to verify the LLM-generated results and incorporate the reasoner’s feedback into our pipeline to ensure logical consistency, thereby generating a hybrid neurosymbolic system. By introducing controlled variations to the background knowledge, we assess the models’ reliance on commonsense versus formal reasoning. Results show that o3-mini achieves near-perfect accuracy across settings, albeit with longer runtime. These findings demonstrate that LLMs have the potential to serve as scalable and flexible DL learners when coupled in a hybrid neurosymbolic setting, offering a promising alternative to symbolic approaches—particularly in contexts where high-quality ontologies are incomplete or unavailable.} }
Endnote
%0 Conference Paper %T Description Logic Concept Learning using Large Language Models %A Adrita Barua %A Pascal Hitzler %B Proceedings of The 19th International Conference on Neurosymbolic Learning and Reasoning %C Proceedings of Machine Learning Research %D 2025 %E Leilani H. Gilpin %E Eleonora Giunchiglia %E Pascal Hitzler %E Emile van Krieken %F pmlr-v284-barua25a %I PMLR %P 160--178 %U https://proceedings.mlr.press/v284/barua25a.html %V 284 %X Recent advances in Large Language Models (LLMs) have drawn interest in their capacity for logical reasoning, an area traditionally dominated by symbolic systems that rely on complete, manually curated knowledge bases represented in formal languages. This paper introduces a framework that leverages pretrained LLMs to generate Description Logic (DL) class expressions from instance-level examples and background knowledge, translated to natural language. The baseline is Concept Induction, a symbolic learning approach that is mostly based on formal logical reasoning over a DL theory. Drawing inspiration from the DL-Learner architecture, our approach replaces traditional symbolic methods with LLM-based models to generate DL class expressions from instance-level data. We evaluate our approach using three benchmark ontologies across two LLMs: gpt-4o and o3-mini. We use a symbolic reasoner, Pellet, to verify the LLM-generated results and incorporate the reasoner’s feedback into our pipeline to ensure logical consistency, thereby generating a hybrid neurosymbolic system. By introducing controlled variations to the background knowledge, we assess the models’ reliance on commonsense versus formal reasoning. Results show that o3-mini achieves near-perfect accuracy across settings, albeit with longer runtime. These findings demonstrate that LLMs have the potential to serve as scalable and flexible DL learners when coupled in a hybrid neurosymbolic setting, offering a promising alternative to symbolic approaches—particularly in contexts where high-quality ontologies are incomplete or unavailable.
APA
Barua, A. & Hitzler, P.. (2025). Description Logic Concept Learning using Large Language Models. Proceedings of The 19th International Conference on Neurosymbolic Learning and Reasoning, in Proceedings of Machine Learning Research 284:160-178 Available from https://proceedings.mlr.press/v284/barua25a.html.

Related Material