LLM-Empowered Knowledge Graph Construction: A Survey

Author: Haonan Bian, Xidian University
Published: ICAIS 2025 (arXiv: 2510.20345)

Core Framing

The paper surveys how LLMs are reshaping knowledge graph construction across three classical stages: ontology engineering, knowledge extraction, and knowledge fusion. The central argument is a paradigm shift from rule-based pipelines to language-driven, generative frameworks.

The most valuable insight is the inversion of the LLM-KG relationship: from “LLMs as tools for building KGs” to “KGs as infrastructure for grounding LLMs.”

Key Contributions

Three bottlenecks of traditional KG construction:

1. Scalability and data sparsity (rule-based systems fail to generalize)
2. Expert dependency and rigidity (schemas require extensive human intervention)
3. Pipeline fragmentation (cumulative error propagation across stages)

Three mechanisms LLMs enable:

1. Generative knowledge modeling (synthesizing structured representations from text)
2. Semantic unification (integrating heterogeneous sources via natural language)
3. Instruction-driven orchestration (coordinating workflows via prompts)

Two paradigms in ontology construction:

• Top-down - LLMs as ontology assistants (semantic modeling, logical consistency)
• Bottom-up - KGs for LLMs (automatic extraction, schema induction, dynamic evolution)

Two paradigms in knowledge extraction:

• Schema-based - emphasizes structure, normalization, consistency
• Schema-free - emphasizes flexibility, adaptability, open discovery

Future Directions Identified

1. KG-based reasoning for LLMs (structured knowledge enhancing inference)
2. Dynamic knowledge memory for agentic systems (KGs as persistent memory substrate)
3. Multimodal KG construction (integrating text, images, audio, video)
4. KGs as “cognitive middle layer” (beyond RAG, providing scaffolding for planning and decision-making)

Connections to heyMHK

• Extends the vectors-vs-graphs conversation with systematic evidence
• Connects to SECI framework: explicit knowledge (KGs) now serves implicit reasoning (LLMs)
• The “dynamic memory substrate” concept echoes the digital garden model
• Schema evolution mirrors the atoms→molecules→organisms maturation pattern

Related: 07-molecule—vectors-vs-graphs, 06-molecule—seci-framework