Bian (2025) — LLM-Empowered Knowledge Graph Construction: A Survey

LLM-Empowered Knowledge Graph Construction: A Survey — Bian (2025)

Bibliographic Reference

Bian, H. (2025). LLM-empowered knowledge graph construction: A survey. In Proceedings of ICAIS 2025. Xidian University. https://arxiv.org/abs/2505.23628

Core Argument

Knowledge graph construction has entered a new paradigm driven by Large Language Models, shifting from rule-based and statistical pipelines toward language-driven, generative frameworks. The paper surveys this shift across the classical three-layer KG pipeline: ontology engineering (top-down LLM-as-assistant and bottom-up KG-for-LLM paradigms), knowledge extraction (schema-based vs. schema-free methods), and knowledge fusion (schema-level, instance-level, and hybrid). Three meta-trends emerge: (1) static schemas are giving way to dynamic schema induction, (2) pipeline modularity is being integrated into generative unification, and (3) symbolic rigidity is yielding to semantic adaptability. KGs are being redefined from static repositories toward “living cognitive infrastructures” that blend language understanding with structured reasoning.

Methods

Narrative survey organizing approximately 30 systems and frameworks from 2023–2025 into a three-layer taxonomy (ontology engineering → knowledge extraction → knowledge fusion), with each layer further split along a schema-based/schema-free or top-down/bottom-up axis. No systematic search methodology, PRISMA flowchart, or inclusion/exclusion criteria are disclosed. The reference list contains approximately 50 entries. The taxonomy is the authors’ qualitative synthesis without formal grounding (e.g., no Delphi method, no inter-rater reliability). Evidence level V — narrative review.

Key Findings

1. Ontology engineering has bifurcated into top-down and bottom-up paradigms. Top-down approaches (Ontogenia, CQbyCQ, NeOn-GPT) use LLMs as intelligent assistants for formal ontology modeling from competency questions or natural language, achieving quality comparable to junior human modelers. Bottom-up approaches (GraphRAG, EDC, AutoSchemaKG) induce schemas from data via clustering and generalization, treating KGs as dynamic memory substrates for LLM reasoning. The field is shifting from “LLMs for Ontology Engineering” toward “Ontologies and KGs for LLMs.”

2. Knowledge extraction operates along a schema-based to schema-free spectrum. Schema-based methods (KARMA, ODKE+) use explicit ontological structures to constrain extraction, evolving from static blueprints toward adaptive, context-aware schema prompting. Schema-free methods (ChatIE, AutoRE, KGGEN) leverage chain-of-thought prompting and instruction tuning to internalize relational structures without predefined templates. The EDC framework bridges both paradigms by coupling open extraction with downstream canonicalization.

3. Knowledge fusion has progressed from ontology-driven to data-driven to LLM-enabled paradigms. Schema-level fusion unifies conceptual backbones via embedding-based clustering and LLM-based deduplication. Instance-level fusion reframes entity alignment as contextual reasoning (LLM-Align, EntGPT) rather than heuristic matching. Hybrid frameworks (KARMA, ODKE+, Graphusion) integrate both levels within unified, prompt-driven workflows — signaling movement toward autonomous, self-evolving KGs.

4. Dynamic schema adaptation is replacing static schema design. Systems like AdaKGC and AutoSchemaKG enable schemas to co-evolve with extracted content, incorporating novel relations and entity types without retraining. Schema-Constrained Dynamic Decoding and Schema-Enriched Prefix Instruction allow schema adaptation at inference time. This represents a conceptual shift from schema guiding extraction to schema co-evolving with it.

5. Future directions center on KGs as cognitive infrastructure. Three trajectories: (a) KG-based reasoning enhancing LLM logical consistency and causal inference, forming a virtuous cycle between construction and reasoning; (b) dynamic KG memory for agentic systems, enabling persistent, time-aware, self-updating knowledge for autonomous agents; (c) multimodal KG construction integrating text, images, audio, and video into unified structured representations.

Concepts Introduced or Used

• Schema-based extraction: Knowledge extraction guided by an explicit ontology or schema providing structural constraints. Ranges from static (fixed ontology → populate) to dynamic (schema co-evolves with extraction).
• Schema-free extraction: Knowledge extraction without predefined ontologies, relying on LLM reasoning patterns (chain-of-thought, instruction tuning) to internalize relational structures. Includes open information extraction (OIE).
• Top-down ontology construction: Ontology designed from competency questions or domain specifications downward; LLMs serve as intelligent assistants to human modelers. Emphasizes semantic consistency and formal rigor.
• Bottom-up ontology construction: Schema induced upward from instance data via clustering, generalization, and canonicalization. Prioritizes coverage and scalability over formal completeness. KGs serve as memory for LLMs.
• Knowledge fusion: Integration of heterogeneous knowledge sources at the schema level (unifying conceptual structure) and instance level (entity alignment, deduplication, conflict resolution).
• Dynamic schema adaptation: Schemas that co-evolve with extracted content, incorporating novel relations and entity types at inference time without retraining (AdaKGC, AutoSchemaKG).
• Competency Questions (CQs): Natural language questions defining the scope and requirements an ontology should satisfy; used as prompts for LLM-driven ontology construction.

Entities Referenced

• Ontogenia — Ontology generation with metacognitive prompting and ontology design patterns (Lippolis et al., 2025)
• EDC (Extract–Define–Canonicalize) — Three-stage pipeline: open extraction → semantic definition → schema canonicalization (Zhang & Soh, 2024)
• AutoSchemaKG — Dynamic schema induction from web-scale corpora via unsupervised clustering (Bai et al., 2025)
• AdaKGC — Schema adaptation via Schema-Enriched Prefix Instruction and Schema-Constrained Dynamic Decoding (Ye et al., 2023)
• GraphRAG — Graph-based retrieval-augmented generation; foundational for bottom-up KG-for-LLM paradigm (Edge et al., 2024)
• KARMA — Multi-agent architecture for schema-guided extraction and fusion (Lu & Wang, 2025)
• ChatIE — Multi-turn dialogue framework for information extraction (Wei et al., 2024)
• Graphusion — Unified prompt-based paradigm for schema-level and instance-level fusion (Yang et al., 2024)

Limitations

• Source type. Narrative survey without systematic search methodology. Evidence level V. No PRISMA flowchart, no disclosed search strategy, no inclusion/exclusion criteria, no inter-rater reliability for taxonomy construction. Cannot assess whether important systems were missed.
• Snapshot of a fast-moving field. Most surveyed systems were published in 2024–2025; many are arXiv preprints. The survey’s shelf life may be short.
• Descriptive, not prescriptive. The paper classifies systems along taxonomic dimensions but does not provide actionable guidance for practitioners choosing between paradigms. No failure mode analysis.
• Missing paradigms. Category-theoretic approaches to knowledge representation (Spivak & Kent, 2012; Patterson, 2017; Lambert & Patterson, 2024) are entirely absent. Wikilink-based lightweight knowledge systems (Obsidian, Roam) are not discussed. The paper equates “knowledge graph” with “RDF/OWL triplestore or property graph.”
• No verification/factuality framework. The paper assumes LLM extraction is correct; there is no treatment of hallucination detection in KG edges, human-in-the-loop verification, or audit trails.
• Aspirational overreach. The conclusion’s claim that surveyed systems enable “autonomous and explainable knowledge-centric AI” overstates: none of the systems are autonomous (all require human schema design or prompt engineering), and LLM reasoning chains are post-hoc rationalizations, not verifiable proofs.

Relevance to Clonal Evolution

This paper has zero biological content. It is a computer science survey about NLP-driven knowledge graph construction. It earns its place in the wiki corpus not through domain contribution but through architectural self-awareness: the wiki itself is a knowledge graph — with typed pages, YAML frontmatter, [[wikilinks]], a tag taxonomy, and lint-enforced integrity constraints — and Bian’s taxonomy provides a diagnostic framework for understanding where our architecture is strong, where it is weak, and what improvements the surveyed systems suggest.

The wiki as a KG in Bian’s taxonomy

Our architecture sits at a specific point in Bian’s framework:

Dimension	Our position	Strength/Weakness
Ontology paradigm	Top-down (expert-designed page types, required sections, tag taxonomy, ologs)	Strong formal structure; inflexible
Extraction paradigm	Schema-based (subagents follow page-type templates)	Consistent output; no open discovery
Fusion paradigm	Manual (cross-domain functors, wikilinks, contradiction registry)	Precise; doesn’t scale
Schema adaptation	Static (CLAUDE.md rules, lint enforcement)	Stable; misses emergent concept types
Knowledge memory	Static (git versioned, no temporal reasoning)	Auditable; no dynamic updating

Three-layer stack with OKF and Lambert (2024)

The Bian survey, the OKF article (okf-local-knowledge-base), and Lambert & Patterson (2024) (lambert2024-double-functorial-semantics) form a complete stack for understanding this wiki’s knowledge architecture:

• OKF specifies the file-format and workflow layer: markdown + YAML frontmatter + [[wikilinks]], the concrete representation this wiki uses.
• Bian (2025) surveys the algorithmic layer: how LLMs populate such formats through schema-guided extraction and fusion. Our ingest pipeline is a manual instantiation of the “static schema-driven extraction” paradigm (§4.1.1).
• Lambert & Patterson (2024) provides the formal semantics layer: typed relationships (arrows, proarrows), commutativity conditions (cells), and structure-preserving cross-domain mappings (double functors) — verifiable correctness criteria the surveyed systems lack.

None of these three papers cites the others. The synthesis — a complete knowledge infrastructure stack from file format through population algorithms to formal verification — is the wiki’s architectural contribution. See docs/superpowers/specs/2026-07-11-kg-construction-stack.md for the full analysis.

Practical relevance

Bian’s survey suggests three improvements to the wiki’s ingest and synthesis infrastructure:

1. Schema-guided extraction from PDFs (§4.1.1). Our subagents write source summaries following page-type templates. Bian’s surveyed systems (KARMA, ODKE+) use ontology snippets — dynamically selected schema subsets — to construct context-aware extraction prompts. This could improve consistency of source summary fields (especially Methods and Key Findings) by providing more targeted schema guidance per paper type.

2. Dynamic tag/schema adaptation (§4.1.2). Our tag taxonomy is manually maintained in CLAUDE.md. AdaKGC’s Schema-Enriched Prefix Instruction mechanism suggests a pattern for detecting when new tags or page types are needed and proposing them for human approval, rather than waiting for lint to catch undefined tags.

3. LLM-enabled cross-domain fusion (§5). Our cross-domain functors (F, G, H) are manually constructed. The surveyed fusion systems (entity alignment via LLM reasoning, schema canonicalization via embedding similarity) suggest lighter-weight approaches for discovering candidate cross-domain mappings that human verification can then confirm or reject.