Reproduce “Heterogeneous LLM Methods for Ontology Learning (Few-Shot Prompting, Ensemble Typing, and Attention-Based Taxonomies)” on OpenResearch

About this paper

Aleksandra Beliaeva, Temurbek Rahmatullaev

A modular LLM-based system for ontology learning, developed by Aleksandra Beliaeva and Temurbek Rahmatullaev, addresses term extraction, typing, and taxonomy discovery without extensive model finetuning. The system achieved multiple top-tier placements across all tasks in the LLMs4OL 2025 challenge, demonstrating competitive performance with efficient prompting and lightweight adaptive modules.

Key insights

• Competitive or top-ranking performance in complex ontology learning tasks can be achieved through lightweight, modular LLM approaches and efficient prompting strategies, without relying on extensive model finetuning.
• Retrieval-augmented prompting consistently improves recall in few-shot term typing by leveraging contextual reasoning from retrieved examples, surpassing embedding-based baselines.
• Even lightweight finetuning (LoRA) of a smaller encoder, when combined with a specialized cross-attention layer, yields substantial improvements in taxonomy discovery over a frozen, larger embedder.

Summary by alphaXiv.