Construction safety standards are in unstructured formats like text and images, complicating their effective use in daily tasks. This paper compares the performance of Retrieval-Augmented Generation (RAG) and fine-tuned Large Language Model (LLM) for the construction safety knowledge retrieval. The RAG model was created by integrating GPT-4 with a knowledge graph derived from construction safety guidelines, while the fine-tuned LLM was fine-tuned using a question-answering dataset derived from the same guidelines. These models' performance is tested through case studies, using accident synopses as a query to generate preventive measurements. The responses were assessed using metrics, including cosine similarity, Euclidean distance, BLEU, and ROUGE scores. It was found that both models outperformed GPT-4, with the RAG model improving by 21.5 % and the fine-tuned LLM by 26 %. The findings highlight the relative strengths and weaknesses of the RAG and fine-tuned LLM approaches in terms of applicability and reliability for safety management.

中文翻译：

---

用于施工安全管理知识检索的检索增强生成与微调大语言模型的性能比较


施工安全标准采用文本和图像等非结构化格式，使其在日常任务中的有效使用变得复杂。本文比较了检索增强生成 （RAG） 和微调大语言模型 （LLM） 在施工安全知识检索中的性能。RAG 模型是通过将 GPT-4 与源自建筑安全指南的知识图谱集成而创建的，而微调的 LLM 是使用来自相同指南的问答数据集进行微调的。这些模型的性能通过案例研究进行测试，使用事故概要作为查询来生成预防性度量。使用指标评估响应，包括余弦相似度、欧几里得距离、 BLEU 和 ROUGE 评分。结果发现，这两个模型的性能都优于 GPT-4，RAG 模型提高了 21.5%，微调后的 LLM 提高了 26%。研究结果突出了 RAG 和微调的 LLM 方法在安全管理的适用性和可靠性方面的相对优势和劣势。