Publications · M.S. thesis, Case Western Reserve University, 2025
Ontology Matching with Knowledge Retrieval and Efficient Large Language Models
TL;DR Retrieval-augmented language models automate building and aligning ontologies, from concept extraction and matching to relation extraction and evaluation. KV-cache quantization lets the models run on limited hardware.
Abstract
Ontologies are formal representations of a domain's concepts and their interrelations. Nonetheless, building rich, accurate ontologies remains a time-consuming, expertise-intensive task. This thesis outlines a comprehensive workflow that integrates retrieval-augmented Large Language Models (LLMs) with structural algorithms to automate the key stages of ontology matching, all while optimizing resource efficiency. First, this work tackles Entity Recognition by surveying contemporary methods for identifying candidate concepts and evaluating LLM performance on the concept extraction task. Next, this thesis introduces Ontology Matching through KROMA, which has the capability to align related concepts efficiently. Then, this addresses Relation Extraction, through detailing an original work that utilize few-shot LLM prompting and minimal fine-tuning to discover hierarchical relationships. After that, this work outlines Ontology Evaluation, drawing on recent studies to summarize key metrics and structural tests for assessing the quality of ontologies. Finally, this discusses how to Optimize LLM for Ontology Matching, focusing on KV cache quantization, so models can run effectively on limited hardware without sacrificing accuracy.
Read
On a phone, open the PDF in your browser's reader.
Cite
@mastersthesis{nguyen2025thesis,
title = {Ontology Matching with Knowledge Retrieval and Efficient Large Language Models},
author = {Nguyen, Lam},
school = {Case Western Reserve University},
year = {2025}
}