Enhancing Human-Machine Interaction through Sentiment Analysis of the Algerian Dialect

Abstract

This thesis investigates sentiment and subjectivity analysis in the Algerian Arabic dialect (Darija) with the overarching aim of enhancing human–machine interaction for dialect speakers. Algerian Arabic presents significant linguistic challenges, including non- standardized spelling, extensive code-switching, and strong regional variation, which com- plicate natural language processing tasks and contribute to the scarcity of reliable com- putational resources. To address these limitations, this work introduces two major contributions. First, we develop and publicly release AlgVec, a suite of word embedding models trained on over 32 million tokens extracted from user-generated content on X (formerly Twitter), covering both Algerian Arabic script and Arabizi. These embeddings are trained using Word2Vec (Skip-gram and CBOW) and FastText architectures and evaluated through intrinsic tasks such as word similarity and nearest-neighbor retrieval, as well as a dialect- focused benchmark adapted from DiaLex. Second, we construct a publicly available corpus of 18,772 manually annotated Algerian dialect tweets labeled into three sentiment classes: positive, negative, and neutral. A dedicated preprocessing pipeline is proposed to handle dialect-specific characteristics, including emoji usage, Arabizi, and code-switching. We conduct a comprehensive benchmark of traditional machine learning approaches (Support Vector Machines and Naive Bayes), deep neural models (Long Short-Term Memory networks and Convolutional Neural Networks), and transformer-based archi- tectures. In addition, we explore a combined embedding strategy that integrates AlgVec with MARBERT for downstream sentiment classification. Experimental results demon- strate that transformer-based models, particularly the MARBERT–LSTM configuration, achieve state-of-the-art performance with an accuracy of 91.23%, significantly outper- forming traditional methods and widely used Modern Standard Arabic embeddings on dialectal tasks. These findings highlight the importance of dialect-aware sentiment analysis for devel- oping more intuitive and culturally aware human–machine interfaces, enabling emotionally intelligent systems such as virtual assistants, customer service platforms, and interactive educational tools to better understand and respond to users’ emotional states expressed in their native dialect.