Abstract
Crane accidents often result in severe injuries and fatalities, necessitating accurate analysis and responsibility prediction to prevent future incidents. However, the scarcity of crane accident-specific datasets and class imbalance pose challenges to developing robust predictive models. This study addresses these challenges by leveraging paraphrasebased data augmentation methodology to expand a dataset of 480 natural language accident descriptions into a balanced dataset of 900 samples. Using Google's Pegasus paraphrasing model, the augmented data improved model performance, as demonstrated by the F1 score increase from 0.29 to 0.5583. Three pre-trained transformer BERT-based models were fine-tuned on the augmented dataset to evaluate their effectiveness in predicting responsible personnel. The results highlight the effectiveness of the paraphrasing technique and transformer-based models in addressing class imbalance and improving classification accuracy. This research demonstrates the potential of natural language text data in enhancing safety analysis and proposes future directions, such as exploring alternative data augmentation methods, leveraging large language models, and experimenting with domain-specific finetuning.

This publication has 10 references indexed in Scilit: