To address the issues of knowledge hallucination and logical conflicts that large language models (LLMs) tend to produce when constructing military ontologies, a multi-agent collaborative ontology development (MACOD) method is proposed. Through role decoupling, the complex knowledge extraction tasks are assigned to 3 professional agents: modeler, logic expert, and tester. The symbolic reasoner is introduced as a strict post-constraint mechanism to check the global consistency of the generated ontology graph. Experiments based on military doctrine data, including comparative and ablation studies, show that compared with single-model approaches, the multi-agent collaborative architecture significantly improves the coverage and accuracy of ontology concept and relationship extraction. At the same time, the symbolic reasoning mechanism effectively identifies and blocks logical contradictions, reducing the system’s hallucination rate.
为克服传统大语言模型在长文本处理中的幻觉,本框架摒弃了缺乏交互的线性流水线设计,构建基于“抽取—校验—反思”闭环博弈(closed-loop game of extraction-validation-reflection)的多智能体协作机制。通过Qwen-Max底座,配合结构化提示词(structured prompt),实现角色能力的逻辑隔离与认知协同。具体运行机制如下。
WangL W, LiR M, YanY, et al. InstructionNER: a multi-task instruction-based generative framework for few-shot NER[PP/OL]. arXiv (2022-03-08)[2026-02-09].
[12]
JoachimiakM P, MillerM A, CaufieldJ H, et al. The artificial intelligence ontology: LLM-assisted construction of AI concept hierarchies[J]. Applied Ontology, 2024,19(4):408-418.
[13]
LippolisA S, SaeedizadeM J, KeskisärkkäR, et al. Ontology generation using large language models[M]//Curry E, Acosta M, Poveda-Villalón M, et al. The semantic web (ESWC 2025. Cham, Switzerland: Springer, 2025:321-341.
ReimersN, GurevychI. Sentence-BERT: sentence embeddings using Siamese BERT-networks[C]//Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP). Stroudsburg, USA: ACL, 2019:3982-3992.
[17]
SchlichtkrullM, KipfT N, BloemP, et al. Modeling relational data with graph convolutional networks[M]//Gangemi A, Navigli R, Vidal M-E, et al. The semantic web (ESWC 2018. Cham, Switzerland: Springer, 2018:593-607.
GlimmB, HorrocksI, MotikB, et al. HermiT: an OWL 2 reasoner[J]. Journal of Automated Reasoning, 2014,53(3):245-269.
[23]
ShinnN, CassanoF, BermanE, et al. Reflexion: language agents with verbal reinforcement learning[PP/OL]. V4. arXiv (2023-10-10)[2026-05-29].
[24]
MadaanA, TandonN, GuptaP, et al. Self-Refine: iterative refinement with self-feedback[C]//Proceedings of the 37th conference on neural information processing systems. Red Hook, USA: Curran Associates Inc., 2023:46534-46594.