Table 5. Overview of the main model types for Arabic hate speech detection, examples, use cases, Arabic-specific adaptations, and key studies in Arabic hate speech detection.

Model type	Example models	Use case	Arabic-specific adaptations	Key studies
Traditional	SVM	Binary classification of hate speech using hand-crafted features.	TF-IDF + Arabic lexicons (e.g., hate word lists).	Alakrot, Murray & Nikolov (2018b)
	Logistic regression	Probabilistic hate speech classification.	Feature engineering for Arabic morphology (e.g., root extraction).	Ousidhoum et al. (2021b)
	Naive Bayes	Lightweight hate speech detection.	Tokenization tailored for Arabic script and stopword removal.	Abozinadah, Mbaziira & Jones (2015)
	Random forest	Ensemble-based classification of offensive language.	Dialect-aware feature selection (e.g., Levantine vs. Gulf Arabic).	Aref et al. (2020)
	k-NN	Nearest-neighbor classification for small datasets.	Normalization of Arabic diacritics and elongations.	Cahyana et al. (2022)
Deep learning	CNN	Character/word-level feature extraction for hate speech.	Character-level embeddings to handle Arabic orthography such as ().	Mohaouchane, Mourhir & Nikolov (2019)
	LSTM	Sequential modeling of hate speech context.	Bidirectional LSTM (BiLSTM) for Arabic morphology and word order.	Al-Ani, Omar & Nafea (2021)
	GRU	Efficient sequential hate speech detection.	Dialect-specific tokenization (e.g., Egyptian Arabic).	Alshalan & Al-Khalifa (2020)
Transformers	AraBERT	Contextual hate speech classification.	Pre-trained on Arabic social media (Twitter) with subword tokenization.	Khezzar, Moursi & Al Aghbari (2023)
	MARBERT	Dialect-aware hate speech detection.	Pre-trained on 1B Arabic tweets covering multiple dialects.	Ben Nessir et al. (2022)
	XLM-Roberta	Cross-lingual hate speech detection.	Fine-tuned on Arabic datasets (e.g., OCA, ArSAS).	Felipe et al. (2022)
Hybrid	CNN-LSTM	Combining spatial and temporal features.	Multi-channel input for Arabic dialects and MSA.	Mohaouchane, Mourhir & Nikolov (2019)
	CNN-GRU	Extract local and sequential features from textual data, GRU captures sequence orders.		Al-Hassan & Al-Dossari (2021)