LLM-based domain-specific knowledge graph construction and hybrid retrieval-augmented generation

Zhou, Xingwei, Li, Shijiang, Liu, Ying ORCID: https://orcid.org/0000-0001-9319-5940, Yang, Wanran, Xiong, Chengyue and Hou, Liang 2026. LLM-based domain-specific knowledge graph construction and hybrid retrieval-augmented generation. Journal of Manufacturing Systems 86 , pp. 1022-1038. 10.1016/j.jmsy.2026.04.023

PDF - Accepted Post-Print Version Available under License Creative Commons Attribution. Download (2MB)

Abstract

Retrieval-augmented generation (RAG) empowered by large language models is emerging as a promising approach for industrial knowledge analysis and decision support. However, in knowledge-intensive industrial domains, existing RAG frameworks often struggle with incomplete or noisy retrieval results, which limits their effectiveness in supporting high-precision industrial applications. To address this challenge, this study proposes an intent-constrained hybrid RAG (IC-HRAG) framework based on an ontology-enhanced domain-specific knowledge graph. The framework constructs a knowledge graph by integrating text classification with knowledge extraction based on large language models, followed by entity disambiguation and graph completion to improve graph quality. An intent-constrained hybrid retrieval algorithm projects queries onto the ontology space, guiding retrieval to balance evidence completeness and relevance. IC-HRAG unifies structured knowledge and unstructured text, enabling high-density, low-noise evidence organization. A case study involving a technical analysis task for superplastic forming equipment was conducted. Experimental results demonstrate that IC-HRAG consistently outperforms various baseline methods, achieving a superior precision of 97.22% and a peak F1 score of 92.72%. Moreover, comprehensive evaluations of decision performance, generation quality, and retrieval evidence quality further confirm its effectiveness, thereby supporting reliable intelligent analysis and decision-making in knowledge-intensive industrial domains.

Item Type:	Article
Date Type:	Publication
Status:	Published
Schools:	Schools > Engineering
Additional Information:	RRS applied
Publisher:	Elsevier
ISSN:	0278-6125
Date of First Compliant Deposit:	15 April 2026
Date of Acceptance:	8 April 2026
Last Modified:	06 May 2026 14:45
URI:	https://orca.cardiff.ac.uk/id/eprint/186397

Actions (repository staff only)

Edit Item