Human Versus AI: A Comparative Study of Zero-Shot LLMs and Transformer Models Against Human Annotations for Arabic Sentiment Analysis

Accurate sentiment analysis in Arabic natural language processing (NLP) remains a complex task due to the language’s rich morphology, syntactic variability, and diverse dialects. Traditional annotation approaches require human experts, face significant challenges related to inter-annotator agreement and dialectal understanding. Recent advances in transformer-based models and large language models (LLMs) offer new techniques to generate annotations. This paper presents a comparative evaluation of three sentiment annotation strategies applied to Saudi dialect tweets: human expert labeling, fine-tuned transformer models (specifically CAMeLBERT-DA), and zero-shot inference using GPT-4o. The selected CAMeLBERT-DA which is already trained specifically for Arabic sentiment tasks and dialects, demonstrates robust performance with fast, scalable predictions. On the other hand, the selected GPT-4o shows competitive zero-shot accuracy without fine-tuning, making it a practical solution for real-time applications. We investigate how each approach performs on two datasets, both of more than 4,000 Saudi tweets covering a wide spectrum of dialects and sentiment expressions. Our methodology involves analyzing consistency across annotations using interrater agreement metrics such as Cohen’s Kappa, Pearson correlation, and class-specific agreement rates. The results reveal that while human annotations capture cultural and context subtleties, they suffer from inconsistency, particularly in ambiguous or dialect-specific cases. This study contributes to the growing body of work on annotation methodologies by highlighting the strengths and limitations of both human and AI-based annotators in Arabic NLP. Our findings suggest that the zero-shot use of domain-specific transformers like CAMeLBERT-DA with general-purpose LLMs such as GPT-4o have a moderate correlation compared to actual human annotators. The paper concludes with recommendations for building reliable ground truth datasets and integrating AI-assisted labeling into Arabic NLP tasks.

@article{Alahmadi2025,
  title     = {Human Versus AI: A Comparative Study of Zero-Shot LLMs and Transformer Models Against Human Annotations for Arabic Sentiment Analysis},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {8},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Dimah Alahmadi},
  doi       = {10.14569/IJACSA.2025.0160882},
  url       = {https://doi.org/10.14569/IJACSA.2025.0160882}
}