Hybrid Embedding and GPT-5 Instruction-Augmented Framework for Arabic Hate Speech Detection
محورهای موضوعی : Natural Language ProcessingAnas Khudhur Abbas al juboori 1 , Heshaam Faili 2
1 - University of Tehran, Department Pardis Alborz, Computer Engineering
2 - Tehran University
کلید واژه: Hate speech, Large language model, Supervised fine tuning, Arabic, ,
چکیده مقاله :
Hate speech refers to expressions that incite violence, hostility, or discrimination against individuals or groups based on attributes such as religion, ethnicity, gender, or national origin. With the rapid growth of social media, hate speech and abusive language have become increasingly widespread, raising concerns about user safety, social cohesion, and responsible online communication. Detecting such content is highly challenging due to the massive scale of user-generated posts, the informal and evolving nature of online dialects, and the limitations of current automated systems. While significant progress has been achieved in high-resource languages such as English, research on Arabic, particularly its dialectal varieties, remains limited despite its global importance. This study addresses these gaps by proposing a comprehensive methodology for hate speech and abusive language detection in Arabic. We adopt hybrid embeddings by combining AraBERT-v2 with Multilingual MPNet to capture both dialect-specific and cross-lingual semantic features. Dimensionality reduction with UMAP and clustering using HDBSCAN are refined iteratively through Silhouette and Coherence scores. From these clusters, linguistic grounded instructions are generated with GPT-5 and evaluated using a two-stage process: LLM-as-a-Judge prompt evaluations for automated assessment, followed by human validation to eliminate ambiguous cases. The refined instructions are then used to fine-tune a classifier on the L-HSAB dataset. The final instruction-augmented model achieved over 92% accuracy, demonstrating that combining hybrid embeddings, adaptive clustering, and instruction-based fine-tuning provides an effective framework for dialectal Arabic hate speech detection and can be extended to other languages.
[1] K. Gasmi et al., "Hybrid Feature and Optimized Deep Learning Model Fusion for Detecting Hateful Arabic Content," IEEE Access, 2025.
[2] K. E. Daouadi, Y. Boualleg, and O. Guehairia, "Systematic investigation of recent pre-trained language model for hate speech detection in arabic tweets," ACM Transactions on Asian and Low-Resource Language Information Processing, 2024.
[3] B. Corporation, "Most Spoken Languages in the World," 2025. [Online]. Available: https://www.berlitz.com/blog/most-spoken-languages-world.
[4] F. Husain and O. Uzuner, "A survey of offensive language detection for the arabic language," ACM Transactions on Asian and Low-Resource Language Information Processing (TALLIP), vol. 20, no. 1, pp. 1-44, 2021.
[5] M. M. Abdelsamie, S. S. Azab, and H. A. Hefny, "The dialects gap: A multi-task learning approach for enhancing hate speech detection in Arabic dialects," Expert Systems with Applications, vol. 295, p. 128584, 2026.
[6] A. Dahou et al., "A Survey on Dialect Arabic Processing and Analysis: Recent Advances and Future Trends," ACM Transactions on Asian and Low-Resource Language Information Processing, 2025.
[7] S. T. Doghmash and M. Saad, "Arabic Hate Speech Identification and Masking in Social Media using Deep Learning Models and Pre-trained Models Fine-tuning," arXiv preprint arXiv:2507.23661, 2025.
[8] S. Salloum, K. Tahat, A. Mansoori, R. Alfaisal, and D. Tahat, "Leveraging K-Means Clustering for Analysis of Arabic Hate Speech Tweets," in Global Congress on Emerging Technologies (GCET-2024), 2024: IEEE, pp. 282-285.
[9] A. H. Ahmed, R.-J. Yew, X. Minocher, and S. Venkatasubramanian, "Navigating Dialectal Bias and Ethical Complexities in Levantine Arabic Hate Speech Detection," arXiv preprint arXiv:2412.10991, 2024.
[10] A. Charfi, M. Besghaier, R. Akasheh, A. Atalla, and W. Zaghouani, "Hate speech detection with ADHAR: a multi-dialectal hate speech corpus in Arabic," (in English), Frontiers in Artificial Intelligence, Review vol. Volume 7 - 2024, 2024-May-30 2024, doi: 10.3389/frai.2024.1391472.
[11] K. E. Daouadi, Y. Boualleg, and K. E. Haouaouchi, "Ensemble of pre-trained language models and data augmentation for hate speech detection from Arabic tweets," arXiv preprint arXiv:2407.02448, 2024.
[12] W. Aldjanabi, A. Dahou, M. A. Al-Qaness, M. A. Elaziz, A. M. Helmi, and R. Damaševičius, "Arabic offensive and hate speech detection using a cross-corpora multi-task learning model," in Informatics, 2021, vol. 8, no. 4: MDPI, p. 69.
[13] W. Antoun, F. Baly, and H. Hajj, "Arabert: Transformer-based model for arabic language understanding," arXiv preprint arXiv:2003.00104, 2020.
[14] N. Reimers and I. Gurevych, "Sentence-bert: Sentence embeddings using siamese bert-networks," arXiv preprint arXiv:1908.10084, 2019.
[15] L. McInnes, J. Healy, and J. Melville, "Umap: Uniform manifold approximation and projection for dimension reduction," arXiv preprint arXiv:1802.03426, 2018.
[16] L. McInnes, J. Healy, and S. Astels, "hdbscan: Hierarchical density based clustering," J. Open Source Softw., vol. 2, no. 11, p. 205, 2017.
[17] H. Mulki, H. Haddad, C. B. Ali, and H. Alshabani, "L-hsab: A levantine twitter dataset for hate speech and abusive language," in Proceedings of the third workshop on abusive language online, 2019, pp. 111-118.
[18] H. Al-Jarrah, M. Al-Smadi, M. Hammad, and F. Shannaq, "Using Deep Learning Techniques to Detect Hate and Abusive Language in Arabic Tweets," International Journal of Intelligent Engineering & Systems, vol. 17, no. 5, 2024.
