Enhancing PII Detection in Student Essays: A Longformer-Based Approach with Synthetic Data Augmentation

This paper investigates the efficacy of Microsoft's Presidio SDK for detecting Personally Identifiable Information (PII) in student essays. It evaluates the performance of a fine-tuned Longformer model as an alternative. While Presidio is widely adopted for PII recognition across domains, we explore its effectiveness within a novel domain-specific dataset comprising 22,000 student essays from a Vanderbilt University MOOC. The dataset's PII annotations include categories such as student names, email addresses, and phone numbers. Our study hypothesizes that a customized Longformer model, fine-tuned on a domain-specific, human-annotated dataset, will significantly outperform Presidio in PII detection, particularly for context-sensitive entities. To address label imbalance in the training data, we augment it with 1,000 synthetically generated essays. The Longformer model's architecture combines local and global attention mechanisms, enabling effective processing of long-form text. Our results demonstrate that the Longformer model trained with synthetic data augmentation achieves superior performance, with an F5 score of 0.936, compared to Presidio's 0.735. These findings underscore the importance of domain-specific datasets and synthetic data augmentation for enhancing PII detection in specialized text corpora.