LLM-Empowered Knowledge Graph Construction: A Comprehensive Survey

Large Language Models are transforming how we build knowledge graphs—shifting construction from brittle, rule-based pipelines toward unified, generative frameworks. This survey maps that transformation across three core stages: ontology engineering, knowledge extraction, and knowledge fusion.

Why Traditional KG Construction Falls Short

Classical KG pipelines suffer three persistent failures. First, rule-based and supervised systems fail to generalize across domains. Second, schema and ontology design demand heavy expert intervention. Third, disjoint pipeline stages propagate errors cumulatively. These limitations block the development of large-scale, self-evolving knowledge graphs.

LLMs address all three by enabling generative knowledge modeling from unstructured text, semantic unification across heterogeneous sources, and instruction-driven orchestration of complex workflows.

Ontology Engineering: Two Complementary Directions

Top-down (LLMs as ontology assistants). LLMs translate natural language specifications—competency questions, user stories, domain descriptions—into formal OWL ontologies. Frameworks like Ontogenia use metacognitive prompting and self-reflection to improve structural consistency. CQbyCQ automates the translation from requirements to ontological schemas. Empirical evaluations show GPT-4 outputs approach the quality of novice human modelers, validating LLMs as active design collaborators rather than passive tools.

Bottom-up (KGs for LLMs). Here the knowledge graph serves as dynamic infrastructure for LLM reasoning rather than a static repository. GraphRAG and OntoRAG generate instance-level graphs from raw text, then abstract schemas through clustering. The EDC (Extract–Define–Canonicalize) framework advances this into a three-stage pipeline: open extraction, semantic definition, and schema normalization. AutoSchemaKG integrates both schema-based and schema-free paradigms into a unified architecture supporting enterprise-scale, real-time KG evolution.

Knowledge Extraction: Schema-Based vs. Schema-Free

Schema-based extraction operates under explicit structural constraints. Early frameworks like KARMA use multi-agent architectures to enforce entity normalization within fixed ontological boundaries. More recent approaches introduce adaptive flexibility: AdaKGC uses Schema-Enriched Prefix Instructions and Schema-Constrained Dynamic Decoding to handle schema drift without retraining. AutoSchemaKG induces schemas dynamically from large corpora, enabling the schema to co-evolve with extracted content.

Schema-free extraction acquires structured knowledge without predefined ontologies. Chain-of-Thought prompting replaces explicit schemas with reasoning-driven organization. AutoRE fine-tunes models on a Relation–Head–Facts pipeline to internalize relational regularities. ChatIE reformulates extraction as multi-turn dialogue. KGGEN decomposes extraction into sequential steps—entity detection then relation generation—to reduce error propagation. Open Information Extraction (OIE) extends this further, discovering all possible triples without type constraints and feeding them into canonicalization pipelines.

Knowledge Fusion: Aligning Heterogeneous Sources

Schema-level fusion unifies structural backbones across knowledge graphs. The field has progressed from ontology-driven consistency (rigid but precise) to data-driven embedding alignment (LKD-KGC) to LLM-enabled canonicalization (EDC), where natural language definitions of schema components enable cross-schema mapping with semantic precision.

Instance-level fusion aligns concrete entities across graphs. KGGEN uses iterative LLM-guided clustering. LLM-Align frames alignment as a constrained multiple-choice reasoning task. EntGPT applies a two-phase pipeline: candidate generation followed by targeted reasoning. COMEM cascades smaller and larger LLMs to balance efficiency and accuracy at scale.

Hybrid frameworks integrate both levels end-to-end. KARMA coordinates specialized agents for schema alignment, conflict resolution, and quality evaluation. Graphusion performs alignment, consolidation, and inference within a single generative cycle—signaling a shift toward autonomous, self-evolving knowledge graphs.

Future Directions

KG-based reasoning. Structured KGs integrated into LLM reasoning pipelines improve logical consistency and causal inference. High-quality KGs enable explainable, verifiable inference—and stronger reasoning in turn supports more robust automated KG construction, creating a self-improving cycle.

Dynamic knowledge memory for agents. LLM agents need persistent memory beyond finite context windows. Frameworks like A-MEM model memory as interconnected notes with contextual metadata. Zep employs temporal knowledge graphs to manage fact validity and support time-aware reasoning. These approaches treat KGs as long-term, interpretable memory enabling continuous learning and multi-agent collaboration.

Multimodal KG construction. VaLiK cascades Vision-Language Models to translate visual features into textual form, then applies cross-modal verification to assemble multimodal KGs without manual annotation. Key challenges remain in modality alignment, scalability, and robustness under missing data.

KGs beyond RAG. Knowledge graphs increasingly serve as cognitive middle layers for querying, planning, and decision-making—not just retrieval backbones. Applications span explainable recommendation (CogER) and biomedical predictive modeling (PKG-LLM).

Conclusion

LLMs reshape KG construction along three clear trajectories: static schemas become dynamic induction, modular pipelines become generative frameworks, and symbolic rigidity becomes semantic adaptability. The result is knowledge graphs that function as living, cognitive infrastructure. Realizing their full potential requires advances in scalability, reliability, and continual adaptation—with multimodal integration and knowledge-grounded reasoning as the key frontiers.