An ELECTRA-Based Advanced Deep Learning Approach for Hate Speech Detection in Turkish Social Media

The proliferation of hate speech on social media platforms, particularly in linguistically rich contexts such as Turkish, has serious psychological impacts on individuals and poses challenges to creating safe online environments. This study presents an innovative approach to detecting hate speech within the Homophobic–Abusive Turkish Comments (HATC) dataset. Our approach develops a classification-focused framework utilizing the dbmdz/electra-base-turkish-cased-discriminator model, which is pre-trained specifically for Turkish. The HATC dataset consists of manually labeled Instagram comments. The model was trained using TensorFlow and the Hugging Face Transformers library for three epochs with the Adam optimizer. Achieving both 96% F1-macro and 96% F1-micro scores, the model surpasses the best average F1 scores reported in the existing literature. This demonstrates the superiority of the Efficiently Learning an Encoder that Classifies Token Replacements Accurately-based deep learning model in effectively capturing the morphological complexity and contextual nuances of Turkish in hate speech classification tasks. By offering a robust language-specific solution for automatic content moderation, this study contributes to digital safety and holds potential for integration into real-time hate speech filtering systems.