LLM-based Automatic Ontology Construction: Empowering Hybrid Intelligent Systems with Structured Memory and Verifiable Reasoning Capabilities

This study addresses the limitations of LLMs in memory and reasoning by proposing a hybrid architecture: combining LLMs with an external ontology memory layer, automatically constructing RDF/OWL knowledge graphs, and supporting SHACL/OWL constraint validation. Experiments show that ontology enhancement significantly improves multi-step reasoning capabilities and achieves a closed-loop process of generation-verification-correction.

Current LLMs rely on pre-trained parameters for memory, which cannot persist new information and have weak structured reasoning; while RAG alleviates knowledge update issues, vector retrieval is a fuzzy match and lacks precise logical deduction capabilities. To address this challenge, researchers propose a hybrid architecture of LLM + external ontology memory layer.

The architecture integrates three memory mechanisms:

1. Parameter Memory: Implicit knowledge from LLM pre-training, enabling fast responses but cannot be updated and has hallucination risks;
2. Vector Memory: Similarity retrieval based on embeddings, suitable for quick access to unstructured information;
3. Ontology Memory: RDF/OWL structured knowledge graphs, supporting precise semantic reasoning and formal verification. Collaboration among the three: Vector memory for candidate retrieval → Ontology memory provides structured knowledge → Parameter memory handles natural language understanding and generation.

The core contribution of the system is the automated ontology construction process:

• Data Ingestion Layer: Supports heterogeneous data sources such as documents (PDF/Word/web pages), APIs, and dialogue logs;
• Knowledge Extraction Layer: LLM-driven entity recognition, relation extraction, normalization processing, and triple generation;
• Validation and Constraint Layer: SHACL rule checks for structure types, OWL reasoning for logical consistency verification, and manual review for key decisions;
• Continuous Update Layer: Incremental fusion of new triples, conflict resolution, and version management to support backtracking.

Validated on the Tower of Hanoi planning task (a multi-step reasoning benchmark):

• Improved Reasoning Capability: Ontology-structured state representation helps the model understand constraints and goals;
• Reduced Error Rate: Formalized knowledge reduces logical errors and avoids invalid search paths;
• Enhanced Interpretability: The reasoning process is mapped to clear paths in the graph, making decisions transparent and auditable.

Hybrid Context Fusion: LLMs receive fused context from vector retrieval results + graph reasoning results + external tool outputs, balancing flexibility and precision; Generation-Verification-Correction Closed Loop: LLM generates candidates → SHACL/OWL validation → feedback for error correction, improving output reliability and consistency.

Application Scenarios: Enterprise knowledge management (integrating knowledge assets), intelligent customer service (precision service), robot control (task planning), scientific research assistance (literature knowledge extraction); Technical Challenges and Solutions:

• Scale Issues: Graph database optimization, distributed storage, query caching;
• Knowledge Conflicts: Confidence weighting, source priority, manual arbitration;
• Ontology Evolution: Version management + incremental updates;
• LLM Integration Cost: Context caching, precomputation of hot queries.

Limitations: Automatic construction accuracy needs improvement (ambiguity/implicit relationships), high cost of complex reasoning, difficulty in cross-domain ontology alignment; Future Directions: Multilingual ontology construction, integration of neural-symbolic reasoning, federated learning for distributed ontology maintenance, multi-modal model collaboration; Conclusion: The hybrid architecture combines the flexibility of neural networks with the precision of symbolic systems, laying the foundation for reliable intelligent systems and promoting the arrival of the era of trustworthy AI.