Deep arab sentiment exploring the analysis of sentiment in arab social discourse

Recently, multimodal sentiment analysis has become a powerful tool for understanding and recognizing human emotions by collecting data from various sources such as text, audio, and images. Unlike traditional methods that rely on a single input type, these systems seek to understand the entirety of a human expression by simultaneously evaluating language, tone of voice, and facial cues. This integrated approach significantly improves emotion recognition accuracy and facilitates deeper and more accurate estimation of individual emotions. Despite significant progress in developing multimodal models and datasets for English, sentiment analysis in Arabic still lags behind. The complexity of Arabic, its complex syntax, numerous dialects, and distinct grammar make sentiment recognition particularly challenging. Furthermore, there is a lack of a comprehensive, large, and well-detailed multimodal dataset for Arabic. This project aims to construct efficient deep learning models that effectively accommodate the linguistic characteristics of Arabic and enhance the accuracy of sentiment analysis. This thesis examines the effectiveness of multimodal sentiment and emotion analysis using the integration of text, audio, and visual media with advanced deep learning frameworks, the research seeks to enhance the effectiveness of sentiment analysis across multiple datasets, specifically the CMU-MOSI, MELD, and Arabic Multimodal Dataset. The study uses sophisticated models, such as the Multimodal Transformer (MULT), in combination with early and late fusion processes to precisely capture and interpret sentiment analysis information from various different inputs, Finally, this thesis presents a comprehensive study aimed at enhancing the accuracy of sentiment and emotion identification, with specific changes applied to the CMU-MOSI, MELD, and Arabic Multimodal Datasets, resulting in substantial advancements in model performance. Utilizing the modified MULT model with DeBERTa, Whisper, and ViT feature extractors on the CMU-MOSI dataset yielded a performance enhancement, elevating accuracy from 80.0% to 84.89%, or a 4.89% gain, For the MELD dataset, the early fusion strategy was particularly effective, increasing the emotion classification accuracy from 67.33 % to 69.89%, an increase of 3.80%, Arabic multimodal dataset, which initially posed challenges due to limited resources and linguistic diversity, adopting transformer-based fusion methods improved accuracy from 63.46% to 72.73% with the MULT model, reflecting a 9.27% improvement. Collectively, these improvements not only improved accuracy but also reduced error rates and misclassifications, confirming the superiority of multimodal fusion and transformer architectures in diverse linguistic and cultural contexts.