Meronymic Ontology Extraction via Large Language Models

TL;DR

The paper tackles the challenge of automatically extracting meronymic (part-whole) product ontologies from unstructured reviews, a task where manual annotation is costly. It introduces a fully automated four-task pipeline (aspect, synset, concept, relation extraction) powered by large language models and evaluated via an LLM-as-judge, comparing against a BERT-based baseline. Key contributions include a novel LLM-centric pipeline, an evaluation framework using LLM judgments, and empirical gains in ontology relevance across multiple Amazon product categories. The work demonstrates the feasibility and potential of LLMs for meronymic ontology extraction, providing a foundation for broader adoption in product domains and beyond, while highlighting trade-offs in computational cost and the need for standardized benchmarks.

Abstract

Ontologies have become essential in today's digital age as a way of organising the vast amount of readily available unstructured text. In providing formal structure to this information, ontologies have immense value and application across various domains, e.g., e-commerce, where countless product listings necessitate proper product organisation. However, the manual construction of these ontologies is a time-consuming, expensive and laborious process. In this paper, we harness the recent advancements in large language models (LLMs) to develop a fully-automated method of extracting product ontologies, in the form of meronymies, from raw review texts. We demonstrate that the ontologies produced by our method surpass an existing, BERT-based baseline when evaluating using an LLM-as-a-judge. Our investigation provides the groundwork for LLMs to be used more generally in (product or otherwise) ontology extraction.

Figures (1)

• Figure 1: Our complete pipeline for extracting meronymic ontologies from product reviews using LLMs, consisting of the four tasks of aspect, synset, concept and relation extraction. In the example shown here, we begin with a review from a customer. The first task comprises the extraction of the relevant aspects from the review, e.g. watch, strap and wristband, with a fine-tuned LLM. The second task then uses a word embedding to extract synset groups, e.g. discarding wristband since it is a synonym of strap.The third task is to select the most relevant of these with an LLM to obtain the concepts.Finally, in the last task, a fine-tuned LLM extracts the relations between concepts, i.e. identifying that strap is part of (in meronymic relation with) watch, resulting in the final product ontology.