A method for constructing a domain-specific large language model for aircraft carrier air-hub operation parsing

Aircraft carrier air-hub operation parsing is a typical domain-specific structured understanding task in carrier-based aviation scenarios. Its source data are primarily derived from unstructured mission logs, command messages, and support reports, and are characterized by strong temporal dependencies, complex spatiotemporal constraints, dense domain-specific terminology, and tight coupling among aircraft status, resource allocation, and operational phases. These characteristics make automatic parsing prone to errors, including omission of critical information, and structural inconsistency. Although cloud-based large language models (LLMs) offer a feasible solution due to their extensive prior knowledge and strong generative capabilities, their high computational cost and deployment requirements limit their applicability in highly resource-constrained carrier-side environments. In contrast, locally deployable LLMs are more suitable for sensitive operational scenarios, yet their domain-specific parsing capability remains insufficient. This study aims to construct a deployable domain-specific large language model for aircraft carrier air-hub operation parsing, with the goal of improving parsing accuracy, reliability, and consistency under local deployment constraints.

We propose a domain-specific large language model construction method for carrier air-hub operation parsing, termed the carrier air-hub operation parsing large model (CAOP). The proposed framework consists of two tightly coupled stages, namely high-dimensional knowledge distillation during training and collaborative verification during inference. In the training stage, a cloud-based teacher model is employed to provide structured supervision across six key information dimensions, including entity recognition, relation extraction, statistical information extraction, predictive information extraction, joint semantic parsing of textual and numerical data, and abnormal event cue extraction. The outputs across these dimensions are further aligned and fused to construct transferable supervision signals, enabling the distillation of domain-specific parsing capabilities from the teacher model into a locally deployable student model via parameter-efficient optimization. In the inference stage, a carrier aviation operation knowledge base and a case-based retrieval mechanism are introduced to provide external evidence and similar historical instances. Based on these resources, collaborative verification is performed on candidate outputs by enforcing both evidence consistency and structural constraints, thereby reducing hallucinations, minimizing redundancy, and correcting inconsistent information fields.

Under the experimental settings adopted in this study, the proposed method improves the average score of the locally deployed Qwen3-32B model on carrier air-hub operation parsing from 13.3 to 18.4, yielding an absolute gain of 5.1 points and a relative improvement of approximately 38.3%. The resulting performance not only surpasses that of the compared cloud-based models but also approaches expert-level evaluation. The improvements are primarily reflected in more complete information field extraction, stronger evidence alignment, and enhanced structural consistency of the final parsing results.

The results demonstrate that the proposed CAOP framework significantly improves the accuracy, reliability, and consistency of carrier air-hub operation parsing under local deployment constraints. It provides a robust and trustworthy structured data foundation for downstream tasks, including automatic scheduling generation, conflict detection, and optimization solving.