Intelligent analysis of ship collision accidents via Low-Rank Adaptation-based fine-tuning of medium-scale Large Language Models
提出一种利用低秩自适应微调中等规模大语言模型的框架，仅用60个标注样本即可从503份双语船舶碰撞报告中高精度提取风险影响因素，并构建贝叶斯网络进行敏感性分析，为海事安全管理提供决策支持。
The rapid advancement of intelligent maritime accident analysis requires processing large-scale, multilingual data across wide geographic regions. However, significant challenges remain in objectively constructing Risk Influencing Factors (RIFs) and ensuring accurate information extraction with limited computational resources. To address these gaps, a framework for intelligent analysis of ship collision accidents based on Low-Rank Adaptation (LoRA) fine-tuning of medium-scale large language models (LLMs) with limited labeled data was proposed. A bilingual dataset comprising 503 ship collision accident reports was established, and the RIF ontology was derived using a Grounded Theory approach. Using 60 labeled samples, models with ≤ 8 B parameters were fine-tuned, achieving an F1 score of 94.11% on the most challenging accident RIF extraction subtask, surpassing base models by 34.82%. Then, the extracted information was transformed into a 1061-row ×24-column training data matrix via a semantic similarity model, enabling construction of a TAN-BN model. Finally, sensitivity analysis was conducted to identify key RIFs, and case studies were performed to evaluate model performance and validate the proposed framework. The research results showed that the proposed approach advances large-scale, cross-lingual intelligent maritime accident report analysis by improving accuracy and efficiency, reducing computational costs, and supporting reliable safety management decisions.