Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LEC-KG: An LLM-Embedding Collaborative Framework for Domain-Specific Knowledge Graph Construction -- A Case Study on SDGs

Yikai Zeng, Yingchao Piao, Jianhui Li

TL;DR

LEC-KG tackles the challenge of constructing domain-specific knowledge graphs from unstructured SDG policy texts by coupling the semantic strengths of LLMs with the structural reasoning of KG embeddings. It introduces a hierarchical SDG ontology, a single-pass hierarchical extraction framework, and an embedding-driven, evidence-grounded feedback loop that iteratively refines extractions and KG representations. The approach demonstrates strong improvements on Chinese SDG reports, particularly for low-frequency relations, achieving a Micro-F1 of 36.8% and markedly better tail performance versus strong LLM baselines. The work shows practical potential for transforming policy documents into validated knowledge graphs, with scalable, bidirectional learning between language understanding and graph structure. Overall, LEC-KG provides a principled, domain-aware pathway to address schema absence, long-tail relations, and unseen entities in KG construction from text.

Abstract

Constructing domain-specific knowledge graphs from unstructured text remains challenging due to heterogeneous entity mentions, long-tail relation distributions, and the absence of standardized schemas. We present LEC-KG, a bidirectional collaborative framework that integrates the semantic understanding of Large Language Models (LLMs) with the structural reasoning of Knowledge Graph Embeddings (KGE). Our approach features three key components: (1) hierarchical coarse-to-fine relation extraction that mitigates long-tail bias, (2) evidence-guided Chain-of-Thought feedback that grounds structural suggestions in source text, and (3) semantic initialization that enables structural validation for unseen entities. The two modules enhance each other iteratively-KGE provides structure-aware feedback to refine LLM extractions, while validated triples progressively improve KGE representations. We evaluate LEC-KG on Chinese Sustainable Development Goal (SDG) reports, demonstrating substantial improvements over LLM baselines, particularly on low-frequency relations. Through iterative refinement, our framework reliably transforms unstructured policy text into validated knowledge graph triples.

LEC-KG: An LLM-Embedding Collaborative Framework for Domain-Specific Knowledge Graph Construction -- A Case Study on SDGs

TL;DR

LEC-KG tackles the challenge of constructing domain-specific knowledge graphs from unstructured SDG policy texts by coupling the semantic strengths of LLMs with the structural reasoning of KG embeddings. It introduces a hierarchical SDG ontology, a single-pass hierarchical extraction framework, and an embedding-driven, evidence-grounded feedback loop that iteratively refines extractions and KG representations. The approach demonstrates strong improvements on Chinese SDG reports, particularly for low-frequency relations, achieving a Micro-F1 of 36.8% and markedly better tail performance versus strong LLM baselines. The work shows practical potential for transforming policy documents into validated knowledge graphs, with scalable, bidirectional learning between language understanding and graph structure. Overall, LEC-KG provides a principled, domain-aware pathway to address schema absence, long-tail relations, and unseen entities in KG construction from text.

Abstract

Constructing domain-specific knowledge graphs from unstructured text remains challenging due to heterogeneous entity mentions, long-tail relation distributions, and the absence of standardized schemas. We present LEC-KG, a bidirectional collaborative framework that integrates the semantic understanding of Large Language Models (LLMs) with the structural reasoning of Knowledge Graph Embeddings (KGE). Our approach features three key components: (1) hierarchical coarse-to-fine relation extraction that mitigates long-tail bias, (2) evidence-guided Chain-of-Thought feedback that grounds structural suggestions in source text, and (3) semantic initialization that enables structural validation for unseen entities. The two modules enhance each other iteratively-KGE provides structure-aware feedback to refine LLM extractions, while validated triples progressively improve KGE representations. We evaluate LEC-KG on Chinese Sustainable Development Goal (SDG) reports, demonstrating substantial improvements over LLM baselines, particularly on low-frequency relations. Through iterative refinement, our framework reliably transforms unstructured policy text into validated knowledge graph triples.
Paper Structure (86 sections, 5 equations, 6 figures, 8 tables, 2 algorithms)

This paper contains 86 sections, 5 equations, 6 figures, 8 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Knowledge Graph Construction from Text
  4. Knowledge Graph Embedding and Completion
  5. Hierarchical Approaches for Relation Extraction
  6. Integrating Large Language Models with Knowledge Graphs
  7. Methodology
  8. Problem Formulation and Framework Overview
  9. Problem Definition
  10. Bidirectional Collaboration Mechanism
  11. Forward Pass.
  12. Backward Pass.
  13. Cold-Start Initialization
  14. Hierarchical SDG Ontology Design
  15. Design Methodology
  16. ...and 71 more sections

Figures (6)

  • Figure 1: Overview of the LEC-KG framework. Solid arrows indicate forward extraction flow; dashed arrows represent feedback channels: red for KGE$\to$LLM refinement and blue for LLM$\to$KGE enhancement.
  • Figure 2: Entity type hierarchy. Solid arrows indicate subtype relationships; dashed arrows indicate meronymic (part-of) relationships.
  • Figure 3: Long-tail distribution of relation types in our annotated dataset (1,758 triples). The top 10 relations account for 68.2% of instances.
  • Figure 4: Tri-partition validation mechanism with dynamic thresholds and diagnostic feedback generation.
  • Figure 5: Bidirectional refinement: Channel 1 provides evidence-guided CoT feedback for LLM re-extraction; Channel 2 incorporates validated triples for KGE model updating.
  • ...and 1 more figures