Extraction of Traffic Crash Locations From Text Narratives: A Zero-Shot Approach Using Large Language Models
Cristian Arteaga, Inhwa Oh, Chang‐Woo Park, JeeWoong Park · 2024
The location of traffic crashes holds significance in understanding geographic crash safety and devising effective safety interventions. In this regard, analysts often conduct manual investigation by reading through lengthy text narratives to extract crash locations and cross-referencing them with geo-coordinates or descriptive location fields in crash reports. Despite being primary practice, it is error-prone and impractical for large datasets. To address these challenges, this study explores the utilization of Large Language Models (LLMs) to process extensive narrative volumes and identify crash locations. Specifically, this study investigates the capability of LLMs to accurately extract crash locations from narratives in a zero-shot fashion, not requiring the need for expensive and time-consuming fine-tuning. For this, we employ a dataset of narratives for crashes that occurred in Colorado in 2020. We evaluated three LLMs of varying sizes in the given location extraction task under deterministic and non-deterministic generation parameters. The findings reveal that even a small-sized LLM can identify crash locations with a 73% similarity to human extraction. These promising results suggest that LLMs could prove to be valuable tools in assisting analysts with the extraction of information from unstructured text data.