Detecting Hate Speech Targeting Protected Groups in Arabic Using Hypothesis Engineering and Zero-Shot Learning with Ground Validation via ChatGPT

Automatic detection of hate speech in low-resource languages presents a persistent challenge in natural language processing, particularly with the rise of toxic discourse on social media platforms. Arabic, characterized by its rich morphology, dialectal variation, and limited annotated datasets, is underrep-resented in hate speech research, especially regarding content targeting marginalized and protected groups. This study proposes a zero-shot learning approach that leverages Natural Language Inference (NLI) models guided by carefully engineered hypotheses in native Arabic to detect hate speech against protected groups, such as women, immigrants, Jews, Black people, transgender individuals, gay people, and people with disabilities. We formulated nine different Arabic hypothesis groups and employed a zero-shot XNLI model with a baseline embedding-based model, incorporating preprocessing techniques on the HateEval Arabic dataset. The results indicate that the XNLI model achieves up to 80% accuracy in detecting targeted hate speech, significantly out-performing baseline models. Furthermore, a real-world validation using GPT-3 via the ChatGPT interface achieved 54% accuracy in zero-shot conversational settings. These findings highlight the importance of hypothesis design and linguistic preprocessing in zero-shot hate speech detection, particularly in low-resource and culturally nuanced languages offering a scalable and culturally aware solution for moderating harmful content in Arabic online spaces.

@article{FathAlalim2025,
  title     = {Detecting Hate Speech Targeting Protected Groups in Arabic Using Hypothesis Engineering and Zero-Shot Learning with Ground Validation via ChatGPT},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {5},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Ahmed FathAlalim and Yongjian Liu and Qing Xie and Alhag Alsayed and Musa Eldow},
  doi       = {10.14569/IJACSA.2025.0160579},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160579}
}