DELG: An Automatic Framework for Constructing Definition-Enriched Lightweight Domain Knowledge Graphs

ACL ARR 2026 January Submission1224 Authors

Keywords: Knowledge Graph, Large Language Model, Question Answering, Information Extraction, Retrieval-Augmented Generation

Abstract:

Knowledge Graphs (KGs) are widely used to represent structured factual knowledge and serve as key resources for reasoning and question answering tasks. When combined with Large Language Models (LLMs), they enhance both interpretability and precision in complex reasoning tasks. However, integrating KGs with LLMs faces two critical challenges: (i) redundancy in triples, which leads to storage inefficiency and increased retrieval latency, and (ii) conceptual ambiguity caused by insufficient entity definitions, which limits reasoning accuracy.

To address these issues, we propose \textbf{DELG} (Definition Enriched Lightweight Graph), a framework designed to construct lightweight and semantically enriched knowledge graphs. DELG includes two key components. The \textit{Redundancy Removing Component} identifies and eliminates both semantic and structural redundancies, thus reducing storage overhead while maintaining factual completeness and improving retrieval efficiency. The \textit{Definition Enrich Component} hierarchically expands entity definitions based on entity complexity, ensuring that concept representations are contextually clear and semantically precise.
We conduct extensive experiments on a medical dataset derived from \textit{Treating Autoimmune Disease with Chinese Medicine} and evaluate DELG on LLM-based KG-QA tasks.
Code and data are available at \url{https://anonymous.4open.science/r/DELG-FD53}.

Paper Type: Long

Research Area: Retrieval-Augmented Language Models

Research Area Keywords: retrieval-augmented generation,applications,knowledge graphs,knowledge base QA

Contribution Types: NLP engineering experiment, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 1224