Text2Onto-Agent: An LLM Agent-Based End-to-End Automated Ontology Construction Method anda Case Study of Green Building Domain Modeling

Source raw items (1)

• Semantic Scholar2026-07-16 23:31:50
  Text2Onto-Agent: An LLM Agent-Based End-to-End Automated Ontology Construction Method anda Case Study of Green Building Domain Modeling

  Domain ontologies play a crucial role in organizing and integrating heterogeneous knowledge in the construction industry, particularly for the modeling and semantic representation of building codes and standards. Existing ontology construction approaches—including manual, semiautomatic, and automated methods—still rely heavily on domain experts, resulting in high development costs, subjective bias, and limited scalability. Recent large language model (LLM)-driven methods have shown promise for end-to-end ontology construction. However, their direct application to iterative ontology modeling remains challenging. LLM calls over long and dynamically evolving contexts often lead to degeneration and hallucination. Moreover, the inherent stochasticity of LLM outputs, together with the accumulation of structural errors across iterations, can progressively undermine the coherence and reliability of the constructed ontology. To address these challenges, we propose Text2Onto-Agent (T2OA), an agent-based, end-to-end ontology construction framework that formulates ontology modeling as a closed-loop reasoning process. T2OA enables an LLM-based agent to coordinate multiple diagnostic reasoning modules while interacting with a graph-based memory that persistently stores the evolving ontology. This memory provides structured and explicit contextual references across iterations, thereby reducing reliance on long-context prompts, mitigating output stochasticity, and supporting proactive error detection and correction during ontology evolution. A case study in the green building domain demonstrates the effectiveness of the proposed framework. T2OA achieves an F1-score of 0.77 for concept disambiguation, as well as prediction accuracies of 85.22% and 90.35% for parent–child and sibling relations, respectively. Furthermore, T2OA significantly outperforms baseline methods in terms of ontology completeness, highlighting its potential for constructing reliable and domain-adaptive ontologies for building codes and standards. The source code is publicly available at https://github.com/pipiyapi/T2OA .