Automatic propaganda detection: identifying propaganda techniques in English and Arabic texts

<p dir="ltr">Propaganda aims to promote particular narratives to achieve specific purposes or advance certain agendas. It can be used to manipulate people’s opinions towards specific policies or ideologies. Therefore, developing systems to identify propaganda techniques in texts can help individuals and organisations combat bias, promote more accurate information, and make informed decisions on what to believe. Combatting propaganda is particularly important in today’s polarised political and ideological climates, where false or misleading information can be used to manipulate public opinion.</p><p dir="ltr">While groundwork exists for propaganda detection in English and Arabic texts, significant challenges persist. This thesis investigates these complexities, focusing on the challenges posed by class imbalance, overlapping spans, annotation errors, and the granularity of class labels. It presents a cross-linguistic, class-level investigation of propaganda detection in English and Arabic texts, leveraging transformer-based models, including BERT, RoBERTa, and AraBERT. It investigates the effectiveness of machine learning techniques such as data augmentation, random truncation, and back-translation in enhancing dataset robustness and mitigating class imbalance.</p><p dir="ltr">This thesis emphasises the importance of adopting balanced evaluation metrics, especially in imbalanced datasets, and advocates sentence-level evaluation approaches as a practical and interpretable alternative to span-based methods. It offers a holistic approach to addressing the challenges in propaganda detection by proposing refined evaluation frameworks and strategies. Additionally, it introduces a refined label set and comprehensive annotation guidelines to reduce ambiguity and improve data consistency and reliability. These contributions lay the groundwork for future research aimed at enhancing propaganda detection and advancing computational methods in this domain. Finally, this thesis establishes a conceptual boundary between propaganda and persuasion, promoting a linguistically and ethically grounded framework to support more consistent labelling and socially responsible model development within sensitive socio-political contexts.</p>