AOCG-LLM+: automatic ontology construction and generation method

Abstract Conceptual ontology construction, a foundational knowledge engineering task, is constrained by traditional methods’ heavy manual reliance and poor complex-domain adaptability. This study proposes the AOCG-LLM+ framework for fully automated conceptual ontology construction, integrating reinforcement learning with human feedback and a dynamically adjusted radius hierarchical clustering algorithm. Three experiments validated the framework via an LLM-calibrated evaluation system combining quantitative metrics and subjective criteria. Results show AOCG-LLM+ outperforms baselines: the RLHF-driven entity-relation extraction module achieves 92.5% entity precision (10.1% points above plain LLMs) and 88.3% relation F1-score (13 points above BERT+GCN), with 20% higher entity recall than shallow statistical learning-based techniques. The clustering algorithm attains state-of-the-art metrics, achieving superior performance on all clustering metrics. Subjective evaluation demonstrates the framework’s universality across different LLMs. This work verifies AOCG-LLM+’s technical advantages, addresses traditional methods’ hierarchical rationality and consistency defects, and provides a reliable fully automated conceptual ontology construction solution.