Presentation
A Multidimensional Accident Analysis Framework Based on Hazard Triangle and Large Language Models
SessionLBR2: Late-Breaking Results
DescriptionFor effective injury prevention, accurate and detailed understanding of accident causation factors is essential. Raw injury data is typically collected in the form of incident narratives at hospitals and accident reports. Due to the unstructured and noisy nature of the narratives with misspellings and improper grammar, it is challenging to automatically extract specific information about the accident causation factors. Manual analysis of narratives is time and resource intensive. Recent advances in large language models (LLMs) have shown promising performances for various natural language processing tasks involving analysis of noisy textual data. We propose a novel framework to extract the key accident causation factors based on Hazard Triangle from the textual incident narratives using LLMs. To evaluate the performance of this framework, we examined its accuracy on a sample of 50 construction related injury reports from January 2015 to February 2021 obtained from the occupational safety and health administration (OSHA’s) website. We used ChatGPT-3.5 LLM and developed prompts based on the concepts of persona-assignment and chain-of-thought to extract the following three Hazard Triangle elements from the narratives: hazardous elements, initiating mechanisms, and threats. The hazard elements for these cases were also manually identified by researchers and compared against the output from LLM framework. The analysis indicated that the LLM framework observed the following Recall values: 70% for hazardous elements, 82% for initiating mechanisms, and 54% for targets/threats. This indicated that the framework can efficiently and reliably extract structured accident causation information from injury narratives thus enabling quicker and effective prevention measures.
Event Type
Late Breaking Results
TimeWednesday, September 11th3pm - 3:10pm MST
LocationFLW Salon B