A hybrid rule-based NLP and machine learning approach for PII detection and anonymization in financial documents

Abstract

Abstract Safeguarding Personally Identifiable Information (PII) in financial documents is essential to prevent data breaches and maintain regulatory compliance. This research presents a scalable hybrid approach that integrates rule-based Natural Language Processing (NLP), Machine Learning (ML) approaches, and a custom Named Entity Recognition (NER) model for the accurate detection and anonymization of Personally Identifiable Information (PII). A varied and accurate synthetic dataset was created to replicate genuine financial document formats, enhancing model training and assessment. The model has attained a precision of 94.7%, a recall of 89.4%, an F1-score of 91.1%, and an overall accuracy of 89.4% on synthetic datasets. Additional validation on actual financial documents, such as audit reports and vendor bills, revealed a consistent performance with an accuracy of 93%. The study utilizes confusion matrices, ROC curves, and precision-recall curves to evaluate the model which further validates the model’s capabilities and generalization ability. The suggested approach provides a robust and efficient solution for protecting sensitive information in operational financial contexts, markedly enhancing current methods for PII protection.

Keywords

• Machine learning
• Natural language processing
• Personally identifiable information
• Data anonymization
• Financial data security