Automatic Ontology Construction Using LLMs as an External Layer of Memory, Verification, and Planning for Hybrid Intelligent Systems
Summary (EN)
A newly posted arXiv paper proposes a hybrid architecture in which large language models are augmented with an external ontological memory layer built from RDF and OWL knowledge representations. The authors argue that current LLM systems are limited by weak long term memory, poor structural persistence, and the difficulty of formally validating outputs when relying only on parametric knowledge or vector retrieval. Their proposed pipeline automatically constructs and updates ontologies from documents, APIs, and dialogue logs, performing entity recognition, relation extraction, normalization, triple generation, and validation through SHACL and OWL constraints. During inference, the model is not used in isolation. Instead, it operates over a combined context that includes vector retrieval, graph based reasoning, and external tools, turning the overall system into a generation, verification, and correction loop. The paper reports experimental observations on planning tasks such as Tower of Hanoi, where ontology augmentation improved multi step reasoning relative to baseline LLM setups. Although the work is early stage, its application orientation is clear: the architecture is framed as infrastructure for enterprise AI, robotics, and agent systems that require persistent knowledge, explainability, and more reliable decisions over time. If the approach proves robust, it would point toward a practical middle ground between pure end to end LLM systems and heavily hand engineered symbolic stacks, with structured memory acting as a durable substrate for verification and planning in production grade agents.
Summary (ZH)
一篇新近发布在 arXiv 的论文提出了一种混合智能架构，用 RDF 与 OWL 构建的外部本体记忆层来增强大语言模型。作者认为，当前纯 LLM 系统在长期记忆、结构化知识保持以及输出结果的形式化验证方面存在明显限制，单靠参数记忆或向量检索很难支撑需要持续推理和可验证性的复杂任务。为此，论文设计了一条自动化流程，可从文档、API 和对话日志中持续构建并更新本体图谱，具体包括实体识别、关系抽取、规范化、三元组生成，以及基于 SHACL 和 OWL 约束的验证步骤。在推理阶段，LLM 并不是孤立运行，而是与向量检索、图谱推理和外部工具共同构成联合上下文，使系统形成“生成, 验证, 修正”的闭环。论文在包括汉诺塔在内的规划任务上报告称，引入本体层后，多步推理表现优于基线 LLM 方案。虽然这项工作仍处于早期阶段，但其应用导向非常明确，作者将其定位为面向企业 AI、机器人和智能代理系统的基础架构，用于支撑持久知识、可解释性以及更稳定可靠的长期决策。如果后续验证结果成立，这类方法可能为纯端到端 LLM 与高度手工构建的符号系统之间提供一条更可落地的中间路径，即以结构化记忆层作为生产级代理系统中的规划与校验基础。
Source