<![CDATA[Digital propaganda on social media has emerged as a critical challenge to democratic stability and national security. Although transformer-based language models have demonstrated promising performance in text classification, their effectiveness for propaganda detection is often constrained by subtle rhetorical manipulation and severe class imbalance, leading to biased predictions toward majority classes. This study addresses these limitations by proposing a fine-tuned RoBERTa-base model integrated with a class-weighted cross-entropy loss to improve the recognition of minority propaganda instances. The model was trained and evaluated on the SemEval-2020 Task 11 sentence-level corpus containing 14,857 annotated sentences, partitioned into 11,886 training samples and 2,971 test samples using stratified sampling. RoBERTa was selected because its robust pre-training strategy and dynamic masking enable more effective contextual representation learning, while class-weighted loss mitigates the adverse effects of the 22.80% to 77.20% class imbalance by assigning weights of 2.19 and 0.65 to the propaganda and non-propaganda classes, respectively. Under identical fine-tuning settings, the proposed model was compared with BERT-base and DistilBERT-base to ensure a fair architectural evaluation. Experimental results demonstrate that RoBERTa achieved the best performance, attaining 93.47% accuracy and a macro-F1 score of 91.82%, outperforming BERT by 1.28 percentage points and DistilBERT by 2.89 percentage points in macro-F1. These findings demonstrate that combining RoBERTa with class-weighted learning provides a robust and practical approach for propaganda detection, supporting the development of automated content moderation and misinformation monitoring systems for social media platforms. Future work will investigate multilingual propaganda detection and fine-grained propaganda technique classification.]]>
Copyrights © 2026
