1. DSpace at Kasdi Merbah University Ouargla
  2. 2. Thèses de Doctorat
  3. 2. Faculté des nouvelles technologies de l’information et de la communication
  4. Département d'informatique et technologie de l'information - Doctorat
Please use this identifier to cite or link to this item: https://dspace.univ-ouargla.dz/jspui/handle/123456789/39871
Full metadata record
DC FieldValueLanguage
dc.contributor.advisorBachir, SAID-
dc.contributor.authorMohammed Elsadiq, BARMATI-
dc.date.accessioned2026-01-11T09:38:14Z-
dc.date.available2026-01-11T09:38:14Z-
dc.date.issued2026-
dc.identifier.urihttps://dspace.univ-ouargla.dz/jspui/handle/123456789/39871-
dc.descriptionArtificial Intelligenceen_US
dc.description.sponsorshipArabic sentiment analysis is a critical task in natural language processing (NLP) that involves identifying and classifying sentiments expressed in Arabic text. Its complexity arises from the language’s rich morphology, widespread dialectal variation, and frequent use of figurative expressions such as sarcasm. Although traditional machine learning models have contributed to the field of NLP, they often fail to capture these linguistic subtleties, limiting their effectiveness in real world applications. This dissertation introduces two novel transfer learning architectures to advance Arabic sentiment analysis: a multimodal framework and a multi-task learning (MTL) framework. The first architecture combines textual representations from pre-trained Arabic transformer models with tabular categorical and numerical features to form a multimodal input. Evaluated on the ArSenTD-Lev dataset, this approach demonstrates that incorporating heterogeneous modalities enhances sentiment classification performance. The second architecture employs an MTL strategy that simultaneously performs sentiment classification, sarcasm detection, and dialect identification using a shared transformer encoder and task-specific decoders. By leveraging shared contextual knowledge and inter-task Interrelatedness, the MTL model enhances generalization and reduces overfitting. Empirical results across benchmark datasets, including ArSarcasm, ArSenTD-Lev, ASTD, and NADI, validate the superior performance of the proposed MTL model over conventional single-task baselines. Collectively, the multimodal and MTL frameworks contribute to a unified and scalable solution for Arabic sentiment analysis. The dissertation concludes by outlining future research directions, including hierarchical task modeling, task specific attention mechanisms, and the direct integration of tabular features into MTL architectures to further enhance task interaction and model interpretability.en_US
dc.language.isoenen_US
dc.publisherKASDI MERBAHUNIVERSITY OUARGLAen_US
dc.subjectArabic sentiment analysisen_US
dc.subjectArabic sarcasm detectionen_US
dc.subjectMulti-task learningen_US
dc.subjectMultimodal dataen_US
dc.subjectmachine learningen_US
dc.subjecttransfer learningen_US
dc.subjectDeep learningen_US
dc.titleA Transfer Learning-Based Approach for Arabic Sentiment Analysisen_US
dc.typeThesisen_US
Appears in Collections:Département d'informatique et technologie de l'information - Doctorat

Files in This Item:
File Description SizeFormat 
Mohammed_Elsadiq_BARMATI_Doctorat.pdf753,17 kBAdobe PDFView/Open
Show simple item record


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.