Sentiment Analysis for E-commerce in the Maghreb: Enhancing Algerian Dialects Classification with BERT

Abstract

E-commerce platforms have become essential in meeting diverse consumer needs rapidly. For instance, Jumia—the largest e-commerce platform in North Africa—receives a high volume of user reviews that reflect a wide range of opinions regarding products. This diversity challenges platform owners striving to offer high-quality products and leaves buyers uncertain about making the best choices. To address these issues, we developed a sentiment analysis framework specifically tailored to the Algerian dialect. Our approach involved constructing a comprehensive database of user reviews categorized into positive, negative, and neutral sentiments. We further enhanced this resource by compiling a specialized dictionary of commonly used Algerian terms and applying GAN-based expansion techniques, as well as translating reviews into English and French to broaden linguistic coverage. To evaluate our method, we implemented two deep learning classifiers: a Deep Neural Network (DNN) and a BERT-based model. Notably, the BERT model achieved its optimal performance at 20 training epochs, with an accuracy of 95.44%, precision of 93.1%, recall of 95.57%, and an F1-score of 94.7%. These results significantly surpassed those obtained using the DNN model, as confirmed by ROC curve analyses and comparative accuracy evaluations. Our findings demonstrate that the integration of advanced NLP techniques with domain-specific language resources markedly enhances sentiment classification, paving the way for more effective analysis systems in e-commerce applications and the broader incorporation of Maghrebi dialects into scientific research.

Keywords: Sentiment Analysis, Maghreb Dialect, Data Augmentation, Lexicon-Based Approach, classification, BERT