Large Language Model-Assisted Large-Scale Construction Safety Data Visualization and Fault Tree Generation with Improved Prompting Strategies

More than 1,000 lives are lost to construction accidents each year. Natural language processing (NLP) and data-driven analysis, including the use of large language models (LLMs), have been applied to identify and mitigate root causes. However, the limited ability of LLMs to manage large data files constrains automation and analysis, particularly for construction professionals without AI expertise. To address this, a user interface was developed to streamline the conversion of natural language accident descriptions into JSON-formatted fault trees and generate clearer and more readable graphs. Generated code was tested for errors, prompt adherence, and visualization accuracy. An iterative improvement process included a second LLM call to remove unsubstantiated or inaccurate data from fault trees. This study enhances the practical use of LLMs for construction safety management and supports improved safety outcomes across the industry.