Multi-Model Deep Learning Approach for Sentiment Analysis of Saudi Dialectal Arabic

Abstract

In today’s digital era, social media platforms have become integral to modern life, providing corporations, governments, and decision-makers with valuable insights derived from large volumes of user-generated content. However, extracting meaningful information from unstructured textual data remains a significant challenge. To address this issue, text classification techniques, particularly sentiment analysis, have been widely adopted to evaluate public opinions and sentiment polarity. While extensive research has focused on English-language data, low-resource languages such as Arabic, characterised by complex morphology and rich dialectal diversity, have received comparatively limited attention. Arabic sentiment analysis is further complicated by linguistic ambiguity and regional variation, which reduce the effectiveness of conventional NLP methods, especially for informal social media content. Moreover, there is a noticeable lack of annotated datasets and comprehensive studies focusing on Arabic dialect sentiment analysis, with the Saudi dialect being particularly underrepresented. This thesis addresses these challenges by focusing on sentiment analysis of Saudi dialect social media content using a range of machine learning (ML) and deep learning (DL) techniques. The research begins with the collection and manual annotation of a dataset related to Saudi education reform from X (formerly Twitter). This is followed by a systematic evaluation of classical ML models using a wide range of feature extraction methods and pre-trained Arabic word embeddings, establishing strong baselines and identifying effective configurations for Arabic sentiment classification. To capture richer semantic and morphological information, the thesis then proposes a hybrid word embedding strategy that integrates pre-trained AraVec and FastText representations at both tweet and word levels, enabling the development of multiple DL architectures for improved sentiment classification. Experimental results demonstrate that different DL models capture complementary sentiment-bearing linguistic patterns. Motivated by this observation, the thesis introduces the CBiR-LR stacked ensemble, which integrates CNN, Bi-LSTM, and RNN base learners with a Logistic Regression meta-classifier to effectively combine their predictions, enhancing robustness and mitigating individual model biases. To complement the ensemble-based CBiR-LR approach, which exploits architectural diversity over static word embeddings, the thesis further investigates a context-aware modelling paradigm. Leveraging the availability of transformer-based Arabic language models, MAR-BiCAtt is proposed to explore the impact of contextualised embeddings and attention mechanisms on dialectal Arabic sentiment analysis. MAR-BiCAtt integrates MARBERT-generated contextual representations with Bi-LSTM, CNN, and attention components to model token-level semantic interactions and selectively focus on sentiment-relevant expressions. This design enables a principled comparison between ensemble learning over static representations and deep contextual modelling, providing insights into their respective performance characteristics and generalisation behaviour across multiple Arabic sentiment datasets. To evaluate robustness and generalisation, additional experiments were conducted on three public Arabic sentiment datasets spanning different domains. The results demonstrate that the proposed approaches achieve consistent performance and improved generalisation across diverse application areas, including education, healthcare, and socio-political sentiment analysis.