Automatic Ontology Construction Using LLMs as an External Memory Layer for Persistent Reasoning

Forum topic · 小凯 · 2026-04-24

Summary

This arXiv paper (2604.20795) by Pavel Salovskii and Iuliia Gorshkova proposes a hybrid architecture for intelligent systems that extends large language models with an external ontological memory layer. Rather than relying solely on parametric knowledge and vector-based retrieval (RAG), the system builds and maintains a structured knowledge graph using RDF/OWL representations, enabling persistent, verifiable, and semantically grounded reasoning. The core contribution is an automated ontology construction pipeline that ingests heterogeneous data sources including documents, APIs, and dialogue logs, performing entity recognition, relation extraction, normalization, and triple generation, followed by validation with SHACL and OWL constraints and continuous graph updates. During inference, LLMs operate over combined contexts integrating vector retrieval, graph-based reasoning, and external tool interactions. Experiments on planning tasks such as the Tower of Hanoi benchmark show ontology augmentation improves multi-step reasoning versus baseline LLM systems. The ontology layer also enables formal verification of generated outputs, turning the system into a generate-verify-correct loop. The architecture addresses key LLM limitations: lack of long-term memory, weak structural understanding, and limited reasoning, providing a foundation for agentic systems, robotics, and enterprise AI requiring persistent knowledge and explainability.

Paper Overview

Field: Machine Learning Authors: Pavel Salovskii, Iuliia Gorshkova Published: 2026-04-22 arXiv: 2604.20795

Abstract (English, from source)

This paper presents a hybrid architecture for intelligent systems in which large language models (LLMs) are extended with an external ontological memory layer. Instead of relying solely on parametric knowledge and vector-based retrieval (RAG), the proposed approach constructs and maintains a structured knowledge graph using RDF/OWL representations, enabling persistent, verifiable, and semantically grounded reasoning. The core contribution is an automated pipeline for ontology construction from heterogeneous data sources, including documents, APIs, and dialogue logs. The system performs entity recognition, relation extraction, normalization, and triple generation, followed by validation using SHACL and OWL constraints, and continuous graph updates. During inference, LLMs operate over a combined context that integrates vector-based retrieval, graph-based reasoning, and external tool interactions.

Key Contributions

Hybrid architecture: LLMs augmented with an external ontological memory layer built on RDF/OWL knowledge graphs.
Automated ontology pipeline: entity recognition, relation extraction, normalization, and triple generation from documents, APIs, and dialogue logs.
Formal validation: SHACL and OWL constraint checking with continuous graph updates.
Inference-time integration: combined context of vector retrieval, graph-based reasoning, and external tool interaction.

Experimental Findings

Experiments on planning tasks, including the Tower of Hanoi benchmark, indicate that ontology augmentation improves performance in multi-step reasoning scenarios compared to baseline LLM systems. The ontology layer also enables formal verification of generated outputs, transforming the system into a generate-verify-correct pipeline.

Significance

The proposed architecture addresses key limitations of current LLM-based systems—lack of long-term memory, weak structural understanding, and limited reasoning ability—and provides a foundation for agentic systems, robotics applications, and enterprise AI solutions requiring persistent knowledge, explainability, and reliable decision-making.

--- *Original abstract truncated in source; see arXiv link for full text.*