Harnessing Generative Pre-Trained Transformers for Construction Accident Prediction with Saliency Visualization
Byunghee Yoo, Jinwoo Kim, Seongeun Park, Changbum R. Ahn, Taekeun Oh
Abstract
Leveraging natural language processing models using a large volume of text data in the construction safety domain offers a unique opportunity to improve understanding of safety accidents and the ability to learn from them. However, little effort has been made to date in regard to utilizing large language models for the prediction of accident types that can help to prevent and manage potential accidents. This research aims to develop a model for predicting the six types of accidents (caught-in-between, cuts, falls, struck-by, trips, and others) by employing transfer learning with a fine-tuned generative pre-trained transformer (GPT). Additionally, to enhance the interpretability of the fine-tuned GPT model, a method for saliency visualization of input text was developed to identify words that significantly impact prediction results. The models were evaluated using a comprehensive dataset comprising 15,000 actual accident records. The results indicate that the suggested model for detecting the six accident types achieves 82% accuracy. Furthermore, it was observed that the proposed saliency visualization method can identify accident precursors from unstructured free-text data of construction accident reports. These results highlight the advancement of the generalization performance of large language processing-based accident prediction models, thereby proactively preventing construction accidents.
Research summary
A fine-tuned GPT model predicts six construction accident types from accident-report text and highlights words that influence its predictions.
Problem
Manual analysis and classification of unstructured construction accident reports is labor-intensive, potentially biased, and difficult to generalize; existing approaches also provide limited interpretability. The study asks whether fine-tuned GPT models can improve accident-type prediction and reveal influential accident precursors in text.
Method
The researchers fine-tuned GPT models using 15,000 actual construction accident records to classify six categories: caught-in-between, cuts, falls, struck-by, trips, and others. They compared GPT-2.0 and GPT-3.0 with TF-IDF, CNN, and BERT baselines, tested unstructured text alone and combined structured and unstructured data, and developed saliency visualization by deleting individual words and measuring the resulting change in prediction accuracy.
Result
The fine-tuned GPT-3.0 model achieved 82.33% overall accuracy and generally outperformed the baseline models and GPT-2.0; unstructured text alone performed better than combined structured and unstructured data. Saliency visualization identified influential accident precursors and contextual words in accident reports, improving interpretability.