Ontology-Based Graph Construction

• Ontology-based graph construction is an approach that leverages formal ontologies to define schemas, semantics, and constraints for knowledge graphs.
• Modern pipelines combine LLM extraction, rule-based reasoning, and clustering to automatically generate and validate KG instances from unstructured data.
• This methodology improves retrieval and reasoning by enhancing interoperability, explainability, and dynamic updates across various domains.

Ontology-based graph construction is an approach for constructing knowledge graphs (KGs) in which an explicit formal ontology governs the graph’s schema, semantics, and constraints. In this paradigm, the ontology defines a set of classes, properties, axioms, and sometimes rules, which together act as the backbone for the organization, enrichment, and interpretation of instance-level data. Recent developments have leveraged LLMs, neural and logic-based algorithms, and semi-automatic pipelines to address the challenges of domain adaptation, schema completeness, evaluation, and dynamic update in constructing and maintaining ontology-based KGs across diverse domains such as enterprise policy, scientific literature, unstructured documents, and open web-scale data.

1. Formalism and Core Principles

Ontology-based graph construction is formally specified by the tuple $O = (C, P, A)$ where $C$ is the set of classes, $P$ the set of properties, and $A$ the set of axioms (including class hierarchies and logical constraints) (Oyewale et al., 1 Feb 2026, Bian, 23 Oct 2025, Qiang, 22 Mar 2026). The graph itself is a triple $G=(I, T)$, where $I$ is the set of instances and $T \subseteq I \times P \times I$ is the set of RDF triples, interpreted according to $O$. Compliance requires that all terms used in the graph are elements of $(C \cup P)$ and that every element of $(C \cup P)$ is used at least once (no “dead” ontology nodes) (Qiang, 22 Mar 2026).

Ontology design separates schema-level (TBox) information from assertional-level (ABox) content. ABox content is generated by mapping individuals and facts to schema elements. The ontology encodes not only taxonomic structure but also domains/ranges, cardinality, role restrictions, and additional logical axioms underpinning validity and semantics.

2. Automated and Semi-Automated Construction Pipelines

Recent pipelines for ontology-based graph construction employ a mixture of LLM-driven extraction, rule-based reasoning, embedding-based clustering, and human-in-the-loop validation. Key paradigms include:

• Extraction Phase: LLMs are used to extract candidate classes and properties from unstructured text, often enforced by explicit schemas (e.g., Pydantic models) to ensure predictable, type-safe output (Oyewale et al., 1 Feb 2026, Kommineni et al., 2024).
• Entailment and Hierarchy Construction: Subsumption relations (subClassOf) among extracted classes are inferred using pairwise entailment queries to LLMs, with confidence scores thresholded (e.g., $C$0) and subsequent transitive reduction to minimize redundant edges (Oyewale et al., 1 Feb 2026).
• Ontology-Aware Chunking and Clustering: In unstructured or semi-structured document settings, hybrid chunking (structural/semantic) segments content, while key entity/relation candidates are clustered (e.g., via name/definition embedding and Leiden community detection) to define classes and relationships (Tiwari et al., 31 May 2025, Ashury-Tahan et al., 2024).
• Ontology-Grounded KG Population: KG triples are instantiated strictly in compliance with the ontology: only properties and classes defined in $C$1 are used as predicates and types. LLM extractors or question-answering steps are schema-constrained for factual/semantic consistency (Feng et al., 2024, Cruz et al., 8 Nov 2025, Yu et al., 1 Dec 2025).

Pipelines such as OntoEKG, OntoMetric, and OntoRAG exemplify the progression from raw/unstructured data to a schema-constrained OWL/RDF representation, with intermediate validation and calibration at both the semantic and structural levels (Oyewale et al., 1 Feb 2026, Yu et al., 1 Dec 2025, Tiwari et al., 31 May 2025).

3. Pattern Matching, Compliance, and Cross-KG Alignment

To ensure rigor and interoperability, ontology-based graphs are evaluated for compliance and alignment with ontological constraints (Qiang, 22 Mar 2026):

• Internal Ontology Compliance: All terms in the KG map to ontology vocabulary; every class/property in the ontology appears in at least one triple. Automated term-matching across four levels (exact, heuristic, semantic, topological) yields a confidence score per term, supporting both alignment and ontology reshaping.
• External Compliance and Alignment: For harmonization of KGs with different ontologies, cross-ontology term matching is performed via embedding and lexico-structural methods, yielding match sets and average match confidence.
• Pattern-Based Compliance: Ontologies (e.g., Brick, BOT, SAREF) are decomposed into fragments/patterns (node/edge templates). Pattern recognition and fragmentation support modular replacement or optimization for coverage, abstraction, and classification performance, with selection guided by minimum-criterion rules (Liebig’s law).

Compliance is quantified by used-entity ratio, matching rate, average confidence, and top-$C$2 recall in matching (Qiang, 22 Mar 2026).

4. Integrating Ontology Construction with Retrieval and Reasoning

Ontology-based graphs serve not only as passive knowledge repositories but as semantic infrastructure for downstream reasoning and retrieval:

• Retrieval-Augmented Generation (RAG): Ontology-backed KG construction outperforms vector-based and flat graph retrieval in terms of answer completeness, factual accuracy, and explainability. Integrating text chunk nodes and schema-based retrievers (e.g., prize-collecting Steiner tree) further enhances retrieval performance (Cruz et al., 8 Nov 2025, Tiwari et al., 31 May 2025, Park et al., 9 Dec 2025).
• Dynamic and Multilevel Retrieval: Multi-level class hierarchies and logical group structures (e.g., hierarchical document trees, k-hop subgraphs) enable focused, ontology-driven retrieval for complex or multi-hop queries (Park et al., 9 Dec 2025, Tiwari et al., 31 May 2025).
• Formal Validation: Two-phase validation protocols combine semantic verification (e.g., type/role sanity checks via LLM) with rule-based schema enforcement (ID uniqueness, required fields, property constraints, relationship cardinality), yielding high semantic accuracy and schema compliance (Yu et al., 1 Dec 2025).

The result is KGs with improved interoperability, explainability, and regulatory auditability, as seen in ESG and industrial standards (Yu et al., 1 Dec 2025, Park et al., 9 Dec 2025).

5. Benchmarking, Evaluation Metrics, and Best Practices

Comprehensive evaluation of ontology-based KG construction covers both structural and instance-level metrics:

Metric	Description	Example Value / Application
Precision/Recall/F1	Standard graph-level triple precision, recall, F1 (exact/fuzzy match)	Fuzzy-match F1=0.724 in OntoEKG (Oyewale et al., 1 Feb 2026)
Used-Entity Ratio	Fraction of ontology terms exercised by the KG	Up to 46.15% after Level 4 matching (Qiang, 22 Mar 2026)
Schema Compliance	Fraction of entities/properties passing schema rules (e.g., VR001–VR006 in OntoMetric)	80-90% in OntoMetric, ~0% baseline (Yu et al., 1 Dec 2025)
Comprehensiveness/Diversity	Coverage and variety in retrieved or generated answers (claim counts/clusters)	88% comprehensiveness vs RAG baseline (Tiwari et al., 31 May 2025)
Human/LLM Judgment	Expert or LLM-based scoring of alignment (0–10, Right/Partial/Wrong)	~79% agreement with human annotation (Kommineni et al., 2024)

Best practices include prompt scaffolding (few-shot, banned individuals), threshold calibration for entailment, logical reasoner integration, human-in-the-loop for intermediate validation, versioned prompt/model/ontology management, and leveraging community vocabularies for schema acceleration (Oyewale et al., 1 Feb 2026, Kommineni et al., 2024, Meckler, 2024).

6. Challenges and Future Directions

Key current challenges and emerging research directions in ontology-based graph construction:

• Hierarchical Reasoning: LLMs and semi-automatic systems may exhibit difficulties in correct directionality for subsumption, imprecise abstraction boundaries, and spurious cycles, requiring external post-processing or reasoner integration (Oyewale et al., 1 Feb 2026).
• Ontology Evolution and Scalability: Robust pipelines must address the need for continual update, self-healing adaptation to new data or schemas without retraining or manual revision, and compliance in federated/multi-KG environments (Bian, 23 Oct 2025, Qiang, 22 Mar 2026).
• Formal Verification at LLM Time Scale: Efficient, lightweight mechanisms for online axiom consistency checking are needed to close the loop between fast schema generation and logical validation (Bian, 23 Oct 2025).
• Benchmark Datasets and Evaluation Standards: The lack of standard, open benchmarks for end-to-end ontology construction and schema-grounded KG population limits reproducibility; current efforts involve domain-specific CQ datasets and hand-labeled gold standards (Park et al., 9 Dec 2025, Oyewale et al., 1 Feb 2026, Feng et al., 2024).
• Multimodal and Mixed-Schema Ontologies: Extension to ontologies spanning text, tabular, visual, and audio content (e.g., scientific diagrams, regulatory tables) presents both modeling and extraction challenges (Park et al., 9 Dec 2025, Bian, 23 Oct 2025).
• Hybrid Workflows: Integrating LLM-driven extraction with post-hoc reasoner-based validation and feedback, possibly employing neural or symbolic consistency checkers, is currently the most robust strategy for production settings (Oyewale et al., 1 Feb 2026, Bian, 23 Oct 2025, Meckler, 2024).

Ontology-based graph construction remains a rapidly evolving area, anchored by formal semantic theory, but increasingly propelled by machine learning, LLMs, and scalable automation. The fusion of automatic extraction, pattern-driven compliance, and rigorous evaluation frameworks ensures KGs that are both usable by downstream systems and trustworthy by design. References: (Oyewale et al., 1 Feb 2026, Bian, 23 Oct 2025, Qiang, 22 Mar 2026, Park et al., 9 Dec 2025, Tiwari et al., 31 May 2025, Yu et al., 1 Dec 2025, Feng et al., 2024, Kommineni et al., 2024, Meckler, 2024, Cruz et al., 8 Nov 2025, Zhapa-Camacho et al., 2023).