PII Detection in Low-Resource Languages Using Explainable Deep Learning Techniques

Safeguarding Personally Identifiable Information (PII) in an increasingly interconnected world presents intimidating challenges, particularly in low-resource languages like Luganda where computational resources for Natural Language Processing (NLP) are scarce. This research attempts to address these challenges, focusing on PII detection in Luganda, a low-resource language spoken in Uganda. The research leverages advanced deep learning methodologies, with attention mechanisms, to enhance PII detection efficacy amidst data scarcity. By directing models to key linguistic features and integrating Explainable AI (XAI) techniques, the study aims to improve both performance and transparency. Three distinct models were implemented i.e. luganda-ner-v6, DeBERTa-v3-Base, and afroxlmr-large-ner-masakhaner. Evaluation results demonstrate promising precision, recall, and F1 scores, while all models perform well, afroxlmr-large-ner- masakhaner consistently excels than the other models on all metrics. for instance he afroxlmr-large-ner-masakhaner model has the highest accuracy with 96.3%, followed closely by luganda-ner-v6 at 95.1%, and deberta-v3-base at 93.9%. The significance of this research extends beyond privacy protection, but also lies in contributing to the broader fields of NLP and privacy technology in low resource languages. This research suggest potential improvement areas including creating PII datasets for multilingual model training, transfer learning, explicit implementation of attention mechanisms, and domain-specific knowledge to advance PII detection and anonymisation in low resource languages.