Résumé:
Sentiment analysis is a core task in Natural Language Processing (NLP), widely applied
in sectors such as digital marketing, politics, and customer feedback systems. However,
its application to Arabic dialects—particularly the Algerian dialect—remains challeng-
ing due to linguistic diversity, lack of standardized orthography, and the prevalence of
non-standard scripts like Arabizi. Existing works are often limited to Modern Standard
Arabic or fail to effectively capture dialect-specific nuances.
To address these limitations, this work proposes a late fusion architecture that com-
bines contextual embeddings from two complementary pre-trained models: DziriBERT,
specialized for Algerian dialect, and MARBERT, trained on a broad set of Arabic di-
alects. The fused representations are then passed to a BiLSTM classifier, selected for
its strong capacity to capture sequential context.
Experiments conducted on two real-world datasets—Djezzy and BrandDZ—demonstrated
the effectiveness of this approach. The BiLSTM-based late fusion model achieved ac-
curacies of 0.83% , respectively, These results confirm that integrating dialectal and
broader Arabic representations, along with appropriate sequence modeling, leads to
more robust and reliable sentiment classification for under-resourced dialects such as
Algerian Arabic.
