LERCause: Deep learning approaches for causal sentence identification from nuclear safety reports

Area of Science:

• Nuclear Safety
• Natural Language Processing
• Machine Learning

Background:

• Identifying causal relationships in nuclear incident reports is crucial for enhancing nuclear safety.
• Automated techniques can improve the accuracy and efficiency of locating and labeling causal sentences.

Purpose of the Study:

• To introduce LERCause, a labeled dataset and methodology for causal sentence classification in nuclear safety.
• To establish a baseline for evaluating and comparing new causal sentence extraction techniques.

Main Methods:

• Utilized three BERT models (BERT, BioBERT, SciBERT) on 10,608 annotated sentences from the Licensee Event Report (LER) corpus.
• Compared BERT models against a keyword-based heuristic, traditional machine learning (Logistic Regression, Gradient Boosting, SVM), and a CNN.

Main Results:

• BERT-centric models significantly outperformed all other tested methods across accuracy, precision, recall, and F1 score.
• BioBERT achieved the highest F1 score of 94.49% through ten-fold cross-validation.

Conclusions:

• BERT-centric models demonstrate superior performance for causal sentence classification in the nuclear safety domain.
• The LERCause dataset and framework provide a reproducible foundation for future research in nuclear safety text analysis.