Arabic Hate Speech Identification and Masking in Social Media using Deep Learning Models and Pre-trained Models Fine-tuning (arXiv)

The identification of hate speech on social media has grown in significance. The current study addresses two issues: 1) identifying hate speech in Arabic text; 2) cleaning the text by substituting star masks that correspond to word length for offending terms. To determine the greatest F1 score for the first challenge, transformer and deep learning models were tested. For the second, text cleaning was viewed as a machine translation issue, in which offensive language is present in the input and is hidden in the output. The detection model’s accuracy was 95% and its Macro F1 score was 92%. The best result for masking outperformed current translation systems, achieving a BLEU score of 0.3 with 1-gram.

https://arxiv.org/abs/2507.23661