面向舰载航空枢纽作业解析的领域大模型构建方法
A Method for Constructing a Domain-Specific Large Language Model for Carrier Air-Hub Operation Parsing
-
摘要: 【目的】舰载航空枢纽作业解析具有专业性强、时空约束复杂等特点，易出现解析结果有误、关键信息遗漏、结构不一致等问题。面向资源高度受限的航空母舰作战场景，本文研究如何构建可部署的领域大模型，提升其在舰载航空枢纽作业解析任务中的准确性、可靠性与一致性。【方法】为此，本文提出面向舰载航空枢纽作业解析的领域大模型构建方法（Carrier Air-Hub Operation Parsing Large Model，CAOP）。训练阶段采用高维蒸馏，将云端教师模型在实体识别、关系识别、统计信息、预测信息、文本-数值联合表意解析、异常事件线索六类关键信息上的解析能力迁移至本地模型；推理阶段引入舰载航空作业知识库与案例检索，采用协同校对对候选输出施加证据约束与结构一致性约束，并修正输出结果。【结果】实验显示，CAOP使本地大模型Qwen3-32B在舰载航空枢纽作业解析任务上的平均得分由13.3提升至18.4（提升5.1分，约38.3%），超越众多云端大模型，并媲美人类专家表现。【结论】实验结果表明，CAOP能在本地可部署条件下显著提升舰载航空枢纽作业解析的准确性、可靠性与结果一致性，为后续调度计划自动生成、冲突检测与优化求解提供可靠数据基础与可信结构化输入。Abstract: Objectives Carrier air hub operation parsing is characterized by strong domain specificity and complex spatiotemporal constraints, and is therefore prone to parsing errors, omission of critical information, and structural inconsistency. For aircraft carrier operational scenarios with highly constrained resources, this study investigates how to construct a deployable domain specific large model to improve accuracy, reliability, and consistency in carrier air hub operation parsing tasks. Methods We propose a domain-specific large model construction method for carrier air-hub operation parsing, termed the Carrier Air-Hub Operation Parsing Large Model (CAOP). In the training stage, high dimensional distillation is employed to transfer the parsing capabilities of a cloud based teacher model across six key information dimensions, namely entity recognition, relation recognition, statistical information, predictive information, text and numerical joint semantic parsing, and abnormal event cue extraction, to a local model. In the inference stage, a carrier aviation operation knowledge base and case retrieval are introduced, and collaborative verification is adopted to impose evidence constraints and structural consistency constraints on candidate outputs and to further refine the parsing results. Results Experiments show that CAOP improves the average score of the local model Qwen3-32B on carrier air-hub operation parsing from 13.3 to 18.4, corresponding to a gain of 5.1 points, approximately 38.3%. CAOP outperforms many cloud-based models and achieves performance comparable to human experts. Conclusions The results demonstrate that CAOP significantly enhances the accuracy, reliability, and consistency of carrier air-hub operation parsing under local deployment constraints. It provides a reliable data foundation and trustworthy structured inputs for subsequent tasks including automatic scheduling plan generation, conflict detection, and optimization solving.
下载: