Sentiment-Annotated Hibapress: A Moroccan News Arabic Dataset (SAHMNAD) predicted using Fine-Tuned Arabic Language Models and Zero-Shot LLMs

This study presents the Sentiment-Annotated Hibapress, a Moroccan News Arabic Dataset (SAHMNAD), a Moroccan news headlines dataset for sentiment analysis, sourced from a national digital news website, Hibapress, with user engagement metrics such as likes and dislikes as sentiment indicators. Unlike prior research focused on social media, this news headline data enables structured sentiment classification in mainly Modern Standard Arabic, addressing linguistic challenges unique to Moroccan digital news.We fine-tune Arabic-specific pretrained language models, including Asafaya-BERT and DarijaBERT, for sentiment classification and compare them with multilingual and general-purpose models. Asafaya-BERT achieves the highest accuracy of 90.74% with an F1 score of 90.58%, significantly outperforming zero-shot large language models such as BLOOM-560m with 78.76% and XLM-R with 66.04%, as well as multilingual models like DistilBERT with 88.65%. The results highlight the effectiveness of fine-tuned models for Moroccan Arabic sentiment classification.This research underscores the necessity of localized Arabic Natural Language Processing (NLP) models and explores large language models for zero-shot sentiment classification. Future directions include integrating reader comments, multimodal sentiment analysis, and cross-lingual approaches to further enhance sentiment classification in Moroccan digital media.