Detection of Personal Data in Unstructured Text Using Deep Learning Models

This thesis explores the detection of personal data in unstructured text using compact deep learning models for Named Entity Recognition (NER). It evaluates the performance of various deep-learning models, including a bidirectional LSTM and multiple versions of BERT, both with and without a Conditional Random Field (CRF) tag decoder. The primary goals are to evaluate these models based on F1-score, precision, recall, and inference speed, and to compare their effectiveness against an existing rule-based algorithm employed by the software company Northern Parklife AB. We find that these models significantly outperform the traditional rule-based approach on our synthetic dataset. Although there are concerns related to generalizability to real-world data, we conclude that integrating deep learning methodologies, possibly by augmenting existing rule-based systems with a compact model as a secondary validation layer, could potentially improve the sensitive personal data detection processes for organizations like Northern.