Možnosti klasifikace extremistického textu pomocí hlubokého učení

Abstract

This thesis explores the application of deep learning models for detecting extremist content, specifically violent threats in online comments. The objective is to design a system capable of maximally efficiently filtering highly imbalanced text data characterized by a rare occurrence of such threatening comments. This system aims to potentially serve as a support tool for security analysts by generating a priority queue that places threats at the top. Within the practical section, a study was conducted using the „threat“ label from the Civil Comments dataset, comparing models ranging from foundational convolutional and recurrent networks to advanced hybrid transformer architectures. The results demonstrate that hybrid transformer architectures achieve the highest utility according to the selected evaluation metrics (AUROC and F1 score). The findings suggest that the synergy between the deep semantic representations of transformers and specialized classification heads is likely crucial for maximizing classification performance under conditions of severe data imbalance.