关键词: 信息抽取, 大语言模型, 模型自我反思, 人机协作, 幻觉

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