PaperlayerView exact revision on arXiv

Source-linked AI summary

LLM-empowered knowledge graph construction: A survey

Haonan Bian

arXiv:2510.20345v1cs.AI

TL;DR

Traditional knowledge graph construction faces scalability and cross-domain generalization challenges. This survey reviews how LLMs reshape ontology engineering, knowledge extraction, and knowledge fusion. It identifies a shift from rule-based, modular pipelines toward unified, adaptive, and generative frameworks.

  • Problem

    Traditional knowledge graph construction faces scalability and cross-domain generalization challenges despite its role in organizing structured knowledge for intelligent applications.

    Source: §1 · p. 1

  • Method

    The survey reviews LLM-driven ontology engineering, knowledge extraction, and knowledge fusion, analyzing methodological paradigms, open challenges, and future directions.

    Source: §1 · p. 2

  • Results

    The survey identifies a shift from rule-based, modular knowledge graph construction toward unified, adaptive, and generative frameworks.

    Source: §7 · p. 9

  • Takeaways & Limitations

    Knowledge graphs serve as dynamic infrastructure providing factual grounding and structured memory for LLMs rather than merely static repositories for human interpretation.

    Source: §3.2 · p. 4

    Traditional ontology engineering frameworks lack flexibility and efficiency for large-scale or continuously evolving knowledge domains, as highlighted by Kotis et al. (2020).

    Source: §2.1 · p. 2

Abstract

from arXiv · show

Knowledge Graphs (KGs) have long served as a fundamental infrastructure for structured knowledge representation and reasoning. With the advent of Large Language Models (LLMs), the construction of KGs has entered a new paradigm-shifting from rule-based and statistical pipelines to language-driven and generative frameworks. This survey provides a comprehensive overview of recent progress in LLM-empowered knowledge graph construction, systematically analyzing how LLMs reshape the classical three-layered pipeline of ontology engineering, knowledge extraction, and knowledge fusion. We first revisit traditional KG methodologies to establish conceptual foundations, and then review emerging LLM-driven approaches from two complementary perspectives: schema-based paradigms, which emphasize structure, normalization, and consistency; and schema-free paradigms, which highlight flexibility, adaptability, and open discovery. Across each stage, we synthesize representative frameworks, analyze their technical mechanisms, and identify their limitations. Finally, the survey outlines key trends and future research directions, including KG-based reasoning for LLMs, dynamic knowledge memory for agentic systems, and multimodal KG construction. Through this systematic review, we aim to clarify the evolving interplay between LLMs and knowledge graphs, bridging symbolic knowledge engineering and neural semantic understanding toward the development of adaptive, explainable, and intelligent knowledge systems.

1 INTRODUCTION

This survey reviews LLM-driven knowledge graph construction across ontology engineering, knowledge extraction, and knowledge fusion. It examines how generative modeling, semantic unification, and instruction-driven orchestration address limitations of conventional construction pipelines.

1 INTRODUCTION

  • Conventional pipelines face scalability and data sparsity, expert-dependent and rigid schemas, and fragmented stages that propagate errors.These limitations hinder self-evolving, large-scale, dynamic knowledge graphs; rule-based and supervised systems often struggle to generalize across domains.Source: §1 · p. 1
  • LLMs support knowledge graph construction through generative knowledge modeling, semantic unification, and instruction-driven orchestration.These mechanisms respectively generate structured representations from text, integrate heterogeneous sources through natural language grounding, and coordinate workflows through prompts.Source: §1 · p. 1
  • The survey synthesizes LLM-driven methods across ontology engineering, knowledge extraction, and knowledge fusion, examining methodological paradigms, open challenges, and future directions.These three stages also form the conventional knowledge graph construction pipeline.Source: §1 · pp. 1–2
  • The review distinguishes top-down and bottom-up ontology construction, schema-based and schema-free extraction, and schema-level, instance-level, and hybrid fusion.Top-down ontology construction treats LLMs as ontology assistants, while the bottom-up perspective considers knowledge graphs for LLMs.Source: §1 · p. 2
  • Future research directions include knowledge graph-based reasoning, dynamic knowledge memory, and multimodal knowledge graph construction.The survey also establishes foundations by reviewing ontology engineering, extraction, and fusion techniques predating LLMs.Source: §1 · p. 2

2 PRELIMINARIES

Traditional knowledge graph construction follows three stages—ontology engineering, knowledge extraction, and knowledge fusion—implemented through rule-based, statistical, and symbolic approaches. These methods face limitations in scalability, generalization, and adaptation to evolving knowledge.

2.1 ONTOLOGY ENGINEERING

  • Traditional ontology engineering formally specifies concepts, relationships, and constraints, emphasizing logical consistency but requiring extensive expert involvement and struggling with scalability and adaptation.Semi-automatic ontology learning sought to derive structures from textual corpora. However, ontology evolution, modular reuse, and dynamic adaptation remained difficult even for advanced frameworks such as NeOn.Source: §2.1 · p. 2

2.2 KNOWLEDGE EXTRACTION

  • Traditional knowledge extraction identifies entities, relations, and attributes, but remains constrained by data scarcity, weak cross-domain generalization, and cumulative error propagation.Interpretable but brittle, domain-specific linguistic rules gave way to statistical and neural approaches, including BiLSTM-CRF and Transformer models. These enabled data-driven feature learning, while dependence on annotated data remained a limitation.Source: §2.2 · pp. 2–3

2.3 KNOWLEDGE FUSION

  • Traditional knowledge fusion integrates heterogeneous sources into consistent graphs, resolving duplication and conflict, but struggles with semantic heterogeneity, large-scale integration, and dynamic updating.Entity alignment determines whether entities across datasets represent the same real-world object. Methods progressed from lexical and structural similarity to shared-space embeddings for scalability and automation, with multi-feature fusion and semantic cues improving alignment robustness.Source: §2.3 · p. 3

3 LLM-ENHANCED ONTOLOGY CONSTRUCTION

LLM-enhanced ontology construction follows two complementary directions: top-down approaches use LLMs to assist formal ontology modeling, while bottom-up approaches induce and evolve schemas to support LLM grounding and reasoning.

3.1 TOP-DOWN ONTOLOGY CONSTRUCTION: LLMS AS ONTOLOGY ASSISTANTS

  • Top-down construction positions LLMs as expert-guided co-modelers, translating natural-language requirements into formal ontologies while emphasizing semantic consistency and structural completeness.Inputs include competency questions, user stories, and domain descriptions; OWL representations provide logical constraints for subsequent extraction and instance population.Source: §3.1 · pp. 3–4, §3.1.2 · p. 4

3.1.1 COMPETENCY QUESTION (CQ)-BASED ONTOLOGY CONSTRUCTION

  • Lippolis et al. (2025b) report autonomous class, property, and logical-axiom generation with consistency comparable to junior human modelers.CQ-based pipelines extend from requirements formalization through validation and instantiation, with experts intervening at critical checkpoints. Ontogenia (Lippolis et al., 2025a) combines metacognitive self-correction with ontology design patterns, while Lippolis et al. (2025b) explore memoryless modular and reflective iterative prompting.Source: §3.1.1 · p. 4

3.1.2 NATURAL LANGUAGE-BASED ONTOLOGY CONSTRUCTION

  • Natural language-based construction induces conceptual hierarchies and relations directly from text, removing dependence on explicitly formulated competency questions.NeOn-GPT (Fathallah et al., 2025) and LLMs4Life (Fathallah et al., 2024) combine prompt-driven workflows, ontology reuse, and adaptive refinement for complex scientific domains. LKD-KGC (Sun et al., 2025) induces open-domain schemas by clustering entity types extracted from document summaries.Source: §3.1.2 · p. 4

3.2 BOTTOM-UP ONTOLOGY SCHEMA CONSTRUCTION: KGS FOR LLMS

  • Bottom-up construction makes KGs dynamic infrastructure for LLM factual grounding and structured memory, prioritizing automatic schema induction and evolution.GraphRAG (Edge et al., 2024) and OntoRAG (Tiwari et al., 2025) extract instance graphs from text, then abstract concepts and relations through clustering and generalization.Source: §3.2 · pp. 4–5
  • EDC (Zhang & Soh, 2024) normalizes induced schemas, while AdaKGC (Ye et al., 2023) incorporates novel relations and entity types without retraining.EDC sequences open extraction, semantic definition, and schema normalization, supporting alignment with existing ontologies or creation of new ones when predefined structures are absent.Source: §3.2 · p. 5
  • AutoSchemaKG (Bai et al., 2025) unifies schema-based and schema-free paradigms for real-time generation and evolution of enterprise-scale knowledge graphs.This deployment-oriented stage prioritizes factual coverage, scalability, and maintainability over purely semantic completeness, treating KGs as external knowledge memory for LLMs.Source: §3.2 · p. 5

4 LLM-DRIVEN KNOWLEDGE EXTRACTION

LLM-driven knowledge extraction follows two paradigms: schema-based methods emphasize normalization and structural consistency, while schema-free methods prioritize adaptability and discovery. These approaches span fixed and evolving schemas, structured generation, and open information extraction.

4.1 SCHEMA-BASED METHODS

  • Schema-based extraction uses explicit schemas for structural guidance and semantic constraints, with approaches evolving from static ontologies toward adaptive frameworks.Its emphasis on normalization, structural consistency, and semantic alignment distinguishes it from schema-free extraction.Source: §4.1 · p. 5, §4 · p. 5

4.1.1 STATIC SCHEMA-DRIVEN EXTRACTION

  • Static schema-driven extraction prioritizes precision, consistency, and interpretability, but rigid ontologies restrict scalability and cross-domain generalization despite increasingly context-aware prompting.Feng et al. (2024) first generate a domain ontology from text, then use it to guide RDF triple extraction. ODKE+ (Khorshidi et al., 2025) selects ontology snippets, while Bhattarai et al. (2024) generates task-specific UMLS prompts; both remain bounded by pre-existing macro-schemas.Source: §4.1.1 · p. 6

4.1.2 DYNAMIC AND ADAPTIVE SCHEMA-BASED EXTRACTION

  • Dynamic schema-based methods let schemas co-evolve with extracted content, supporting adaptive schema learning and laying foundations for continual, self-updating knowledge graphs.AutoSchemaKG (Bai et al., 2025) combines unsupervised clustering, relation discovery, and relation-specific prompts to induce and iteratively evolve schemas. AdaKGC (Ye et al., 2023) addresses schema drift using Schema-Enriched Prefix Instruction and Schema-Constrained Dynamic Decoding, enabling adaptation without retraining.Source: §4.1.2 · p. 6

4.2 SCHEMA-FREE METHODS

  • Schema-free extraction derives structured knowledge without predefined ontologies or relation schemas through two trajectories: structured generative extraction and open information extraction.LLMs identify entities and relations through prompt engineering, instruction tuning, and self-organizing reasoning.Source: §4.2 · p. 6

4.2.1 STRUCTURED GENERATIVE EXTRACTION

  • Structured generative extraction organizes latent relational structures without explicit schemas through guided reasoning, modular prompting, retrieval, and interactive refinement.Nie et al. (2024) uses Chain-of-Thought reasoning, while AutoRE (Xue et al., 2024) internalizes relational regularities through instruction fine-tuning with a Relation–Head–Facts pipeline. Papaluca et al. (2024) retrieves related exemplars to improve factual precision; ChatIE (Wei et al., 2024) refines candidates through multiturn questions; KGGEN (Mo et al., 2025) separates entity detection from relation generation to mitigate error propagation.Source: §4.2.1 · p. 7

4.2.2 OPEN INFORMATION EXTRACTION (OIE)

  • Open information extraction discovers triples without predefined types, prioritizing coverage and discovery over structural regularity and connecting extraction with emergent ontological organization.EDC (Zhang & Soh, 2024) uses few-shot prompting to extract natural-language triples into a raw open knowledge graph, followed by definition and canonicalization.Source: §4.2.2 · p. 7

5 LLM-POWERED KNOWLEDGE FUSION

LLM-powered knowledge fusion aligns schemas, reconciles instances, or integrates both in hybrid frameworks. These approaches increasingly emphasize semantic reasoning and integrated, generative workflows.

5.1 SCHEMA-LEVEL FUSION

  • EDC (Zhang & Soh, 2024) canonicalizes schemas using LLM-generated component definitions and vector similarity, supporting self-alignment and cross-schema mapping.The approach increases automation and semantic precision by comparing natural-language definitions of schema components.Source: §5.1 · p. 7
  • Schema-level fusion has progressed from ontology-driven constraints through data-driven integration to flexible, LLM-enabled semantic canonicalization.Kommineni et al. (2024) constrain extracted triples using predefined ontologies, providing semantic consistency but limited cross-domain flexibility. LKD-KGC (Sun et al., 2025) dynamically merges equivalent entity types through embedding clustering and LLM-based deduplication.Source: §5.1 · pp. 7–8

5.2 INSTANCE-LEVEL FUSION

  • Instance-level fusion increasingly uses contextual reasoning to address entity alignment, disambiguation, deduplication, and conflict resolution.KGGEN (Mo et al., 2025) follows progressive triple extraction with iterative LLM-guided semantic clustering of equivalent entities and relations. LLM-Align (Chen et al., 2024) frames alignment as constrained multiple-choice reasoning, while EntGPT (Ding et al., 2025) generates candidates before targeted reasoning selects the final entity.Source: §5.2 · p. 8
  • Structural retrieval and hierarchical LLM pipelines improve instance disambiguation robustness and large-scale fusion efficiency.Pons et al. (2025) use RAG with class–subclass hierarchies and entity descriptions for zero-shot disambiguation of unseen or ambiguous entities. COMEM (Wang et al., 2024) cascades smaller and larger LLMs, combining lightweight filtering with fine-grained reasoning to improve efficiency while maintaining high semantic accuracy.Source: §5.2 · p. 8

5.3 COMPREHENSIVE AND HYBRID FRAMEWORKS

  • Hybrid frameworks integrate schema and instance fusion into adaptive, generative workflows aimed at autonomous, self-evolving knowledge graphs.KARMA (Lu & Wang, 2025) assigns schema alignment, conflict resolution, and quality evaluation to specialized collaborating agents; ODKE+ (Khorshidi et al., 2025) couples ontology-guided schema supervision with instance-level corroboration. Graphusion (Yang et al., 2024) performs alignment, consolidation, and inference within a single prompt-based generative cycle.Source: §5.3 · p. 8

6 FUTURE APPLICATIONS

Future applications position KGs as foundations for LLM reasoning, persistent agent memory, multimodal knowledge integration, and structured support for querying, planning, and decision-making.

6.1 KNOWLEDGE GRAPH-BASED REASONING FOR LLMS

  • Integrating structured KGs into LLM reasoning aims to improve logical consistency, causal inference, and interpretability, while stronger reasoning could reciprocally automate KG construction.High-quality, well-structured KGs underpin explainable, verifiable inference. KG-based Random-Walk Reasoning (Kim et al., 2024) and KG-RAR (Wu et al., 2025) demonstrate this paradigm’s potential.Source: §6.1 · p. 9

6.2 DYNAMIC KNOWLEDGE MEMORY FOR AGENTIC SYSTEMS

  • Dynamic KGs offer persistent, interpretable agent memory beyond finite context windows, supporting continuous learning, multi-agent collaboration, and self-reflective reasoning.A-MEM (Xu et al., 2025) continually reorganizes interconnected notes enriched with contextual metadata, while Zep (Rasmussen et al., 2025) uses temporal KGs for fact validity and time-aware updates. Future priorities include scalability, temporal coherence, and multimodal integration.Source: §6.2 · p. 9

6.3 MULTIMODAL KNOWLEDGE GRAPH CONSTRUCTION

  • Multimodal KGs unify text, images, audio, and video for richer reasoning, but face modality heterogeneity, alignment noise, scalability, and missing or imbalanced modalities.VaLiK (Liu et al., 2025) cascades pretrained vision-language models to translate visual features into text, then verifies cross-modal correspondence to filter noise and link entities with images without manual annotation. KG-MRI uses multimodal embeddings and contrastive objectives to align modalities in coherent semantic spaces.Source: §6.3 · p. 9

6.4 NEW ROLES FOR KGS IN LLM APPLICATIONS: BEYOND RAG

  • Beyond RAG retrieval, KGs are envisioned as a cognitive middle layer supporting querying, planning, and decision-making for more grounded, interpretable generation.CogER combines intuitive and path-based KG inference for explainable recommendation. PKG-LLM uses domain KGs for knowledge augmentation and predictive modeling in mental health diagnostics.Source: §6.4 · p. 9

7 CONCLUSION

LLMs are shifting knowledge graph construction from rule-based, modular pipelines toward unified, adaptive, generative frameworks. Scalability, reliability, and continual adaptation remain challenges, with future advances centered on prompt design, multimodal integration, and knowledge-grounded reasoning.

7 CONCLUSION

  • LLMs are transforming ontology engineering, knowledge extraction, and knowledge fusion, shifting rule-based, modular pipelines toward unified, adaptive, generative frameworks.The survey examines this transformation across all three stages of knowledge graph construction.Source: §7 · p. 9
  • Three trends emerge: static schemas evolve toward dynamic induction, pipeline modularity becomes generative unification, and symbolic rigidity gives way to semantic adaptability.Together, these shifts blend language understanding with structured reasoning in knowledge graphs.Source: §7 · p. 10
  • Scalability, reliability, and continual adaptation remain challenges; advances in prompt design, multimodal integration, and knowledge-grounded reasoning are key to autonomous, explainable knowledge-centric AI.Source: §7 · p. 10

Related papers

Loading 2510.20345v1…