Abstract: In the aviation domain, stable communication between airborne and ground units requires that unstructured text messages be transformed into structured information in a fixed format for reliable transmission. Although existing approaches have leveraged the instruction-following capabilities of large language models (LLMs) to perform such transformation, the inherent hallucination and lack of interpretability of LLMs—which pose serious threats to information security in safety-critical aviation contexts—have been largely overlooked. To address this issue, a Message Consistency Parsing Framework for the Aviation domain (MCPFA) was proposed, integrating model self-reflection and human–machine collaboration to systematically ensure semantic fidelity and operational safety during the parsing process. First, a consistency verification method based on model self-reflection (CVMSR) was developed, establishing a progressive validation mechanism across three linguistic dimensions—syntax, semantics, and pragmatics—to enable dynamic verification and correction of the structured output. Second, an intelligent aviation message human–machine collaboration inspection system (IAMHCIS) was designed, which uses confidence-aware monitoring and cross-layer consistency analysis to trigger expert intervention precisely and incorporates human feedback into model refinement through a closed-loop knowledge distillation process. Finally, the framework was evaluated through comprehensive experiments. Experimental results demonstrated that MCPFA significantly outperformed existing approaches in both consistency judgment and error correction, effectively identifying and rectifying critical errors caused by term misinterpretation, unit confusion, or violations of aviation regulations. Furthermore, the framework exhibited strong generalization capability on unseen message types. Importantly, continuous performance gains were observed as expert knowledge was integrated, confirming the framework’s adaptability and practical value. The proposed approach provides a viable pathway for the safe and reliable deployment of large language models in aviation and other safety-critical domains, where accuracy, security, and interpretability must be rigorously maintained.

Key words: information extraction, Large Language Model (LLM), model self-reflection, human-machine collaboration, hallucination

摘要： 在航空领域，为确保空中单位与地面单位之间通信的稳定性，非结构化的文本消息需被转化为固定格式的结构化信息数据进行传输。现有方法虽利用大语言模型（LLM）的指令遵循能力实现文本到结构化数据的转换，却普遍忽视了LLM固有的幻觉问题与不可解释性对航空信息安全构成的严重威胁。为此，提出面向航空领域的消息一致性解析框架（MCPFA），通过融合模型自我反思与人机协同机制，系统性保障信息转化过程中的语义保真与操作安全。首先，提出基于LLM自我反思的一致性检验方法（CVMSR），从语法、语义、语用3个维度构建递进式验证机制，实现对结构化输出的动态校验与修正；其次，设计智能化航空消息人机协作检验系统（IAMHCIS），通过置信度感知与跨层一致性分析精准触发专家干预，并将人类反馈融入模型迭代，形成闭环优化；最后，通过系统性实验全面评估框架性能。实验结果表明，MCPFA在一致性判断与修正任务上显著优于现有方法，不仅能有效识别并修复因术语混淆、单位误判或规章违背导致的关键错误，还在未见过的消息类型上展现出良好的泛化能力。更重要的是，引入专家知识后，模型性能持续提升，验证了框架的可成长性与工程实用价值,为航空等安全攸关领域的大模型应用提供了兼顾准确性、安全性与可解释性的新路径。

关键词: 信息抽取, 大语言模型, 模型自我反思, 人机协作, 幻觉