Textual data transformations using natural language processing for risk assessment.

Area of Science:

• Engineering
• Computer Science
• Risk Management

Background:

• Unstructured text in failure reports contains valuable causation insights.
• Existing methods lack solutions for utilizing free-text data in risk assessment.
• Textual data is a crucial, yet underutilized, resource for developing risk assessment methodologies.

Purpose of the Study:

• To bridge the knowledge gap in extracting features from textual data for cause-effect scenario development.
• To create a methodology for risk assessment using unstructured accident report data with minimal manual interpretation.

Main Methods:

• Application of natural language processing (NLP) and text-mining techniques to extract features from accident reports.
• Transformation of extracted features into parametric form using fuzzy set theory.
• Utilization of fuzzy set theory outputs as prior probabilities in Bayesian networks for risk assessment.

Main Results:

• Demonstrated application on microbiologically influenced corrosion incident reports from the Pipeline and Hazardous Material Safety Administration database.
• Validation of a trained Named Entity Recognition (NER) model on eight incidents, showing promising results in feature identification.
• Confirmation of the NER method's robustness and applicability in extracting relevant features from textual data.

Conclusions:

• The proposed methodology effectively extracts features from unstructured text for risk assessment.
• This approach enables the development of cause-effect scenarios for analyzing, predicting, and preventing future mishaps.
• The methodology enhances overall process safety through domain-specific risk assessment.