Designing Sensitive Personal Information Detection and Classification Model for Amharic Text

Abstract

Sensitive information is a classified type of content that should not be disclosed to the public and that can harm the owner of the information if it is disclosed. To protect disclose of sensitive information first, it requires detecting the availability of sensitive information and its domain classification for further analysis. To the best of our knowledge, there is no work attempted for Amharic texts. Models developed for another language cannot be used for Amharic texts language because of morphology, grammar and semantics differences. To address these gaps, we have proposed a model for detecting and classifying personal sensitive information for Amharic texts. We have experimented with three deep learning algorithms: Long Short-Term Memory (LSTM), Bidirectional Long Short-Term Memory (BI-LSTM) and Convolutional Neural Network (CNN) using 7.31K and 6.697K Amharic sentences for sensitivity detection and domain classification respectively. The accuracy of LSTM, BI-LSTM and CNN was 82%, 90% and 87% respectively for sensitivity classification and 88, 93, 89 respectively for domain classification.