Sentiment Analysis of Algerian Darija Comments on Social Media

Abstract

(NLP), especially for understanding public opinions expressed on social media. However, Algerian Arabic remains a low-resource dialect due to the lack of annotated corpora, linguistic tools, and standardized writing forms. This work focuses on sentiment analysis of Algerian Darija comments collected from social media platforms, mainly YouTube and Twitter. In this study, we built and prepared the AlgerianPulse corpus through several stages, including data collection, cleaning, preprocessing, annotation, dataset integration, and balancing. The comments were classified into three sentiment categories: positive, negative, and neutral. Different approaches were investigated, including traditional machine learning models, word embedding techniques such as Word2Vec and FastText, deep learning models, pretrained transformer-based language models, and a Hard Majority Voting ensemble. The experimental results show that pretrained transformer models outperformed the traditional machine learning and deep learning approaches, with DziriBERT achieving the best individual performance. Furthermore, the proposed Hard Majority Voting ensemble achieved the highest overall performance, reaching an accuracy of 86.42% and demonstrating the effectiveness of combining multiple classifiers to improve the robustness and reliability of sentiment prediction. Overall, this work contributes to the development of sentiment analysis resources and models for Algerian Arabic. It highlights the challenges of dialectal Arabic processing, particularly spelling variation, code-switching, informal expressions, and the scarcity of labeled data, while proposing a practical methodology for building and evaluating sentiment classification systems in a low-resource language context.