# Ontology Matching with Large Language Models and Prioritized Depth-First Search

**Authors:** Maria Taboada, Diego Martinez, Mohammed Arideh, Rosa Mosquera  
**arXiv:** [2501.11441](https://arxiv.org/abs/2501.11441)  
**Date:** 2025-01-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2501.11441v2/

## Key points

- Ontology matching (OM) is crucial for data interoperability and knowledge sharing.
- Challenges in OM include the need for large training datasets and limited vocabulary processing in machine learning approaches.
- Recent advancements in Large Language Models (LLMs) have shown promise in OM through a retrieve-then-prompt pipeline.
- A novel approach called MILA embeds a retrieve-identify-prompt pipeline within a prioritized depth-first search (PDFS) strategy to efficiently identify semantic correspondences with high accuracy.
- MILA achieved the highest F-Measure in four out of five unsupervised tasks, outperforming existing OM systems by up to 17% and performing comparably or better than leading supervised OM systems.
- MILA exhibited task-agnostic performance across all tasks and settings while significantly reducing LLM requests.

## Summary

Ontology matching (OM) is crucial for data interoperability and knowledge sharing. However, it faces challenges such as the need for large training datasets and limited vocabulary processing in machine learning approaches. Recent advancements in Large Language Models (LLMs) have shown promise in OM through a retrieve-then-prompt pipeline. Despite this progress, these systems still have limitations in performance and computational overhead. To address these issues, a novel approach called MILA has been introduced. MILA embeds a retrieve-identify-prompt pipeline within a prioritized depth-first search (PDFS) strategy to efficiently identify semantic correspondences with high accuracy. This approach limits LLM requests to only the most borderline cases, reducing computational burden. The effectiveness of MILA was evaluated using the biomedical challenge from the Ontology Alignment Evaluation Initiative. Results showed that MILA achieved the highest F-Measure in four out of five unsupervised tasks, outperforming existing OM systems by up to 17%. It also performed comparably or better than leading supervised OM systems. Additionally, MILA exhibited task-agnostic performance across all tasks and settings while significantly reducing LLM requests. These findings emphasize that high-performance LLM-based OM can be achieved through a combination of programmed (PDFS), learned (embedding vectors), and prompting-based heuristics without relying on domain-specific heuristics or fine-tuning. The authors acknowledge support from the University of Santiago de Compostela and projects AF4EU and SUS-SOIL funded by the European Union's Horizon Europe program. They also express gratitude to Diego Martinez-Taboada for insightful conversations. For more information on MILA, data availability can be found at https://github.com/mariatab/MILA. Additionally, a detailed overview of state-of-the-art OM systems used for comparison in the study is provided in Table A.10 in the appendix.