Hybrid RNN-LSTM Architecture for Sentiment Analysis of Algerian Dialectal Social Media Content

Sentiment analysis in under-resourced dialects like Algerian Arabic (Darija) presents unique challenges due to code-switching, informal orthography, and cultural-linguistic nuances. This study addresses the binary sentiment classification task using a hybrid Recurrent Neural Network–Long Short-Term Memory (RNN-LSTM) architecture, designed to effectively capture sequential dependencies and long-term contextual information. The model is trained on DZSentiA, a curated dataset of annotated Algerian dialect social media posts, and achieves strong performance with an accuracy of 84.7%, an F1-score of 84.45%, a recall of 82,75%, and a precision of 84.7%. These results surpass several baseline methods, highlighting the potential of deep learning approaches in low-resource dialectal settings. This work contributes to dialect-specific Natural Language Processing (NLP) by demonstrating the feasibility and effectiveness of deep models in sentiment detection for Algerian Darija, and supports the broader goal of developing culturally aware tools for online discourse analysis.