This paper presents a CLIP-based transfer learning approach for classifying historical archive images in the Nakba Image Classification Shared Task at the Nakba-NLP 2026 Workshop (LREC 2026). The task involves distinguishing images depicting destroyed or damaged infrastructure from those showing intact scenes using a dataset of 2,001 images collected from Instagram posts published by Palestinian content creators and journalists in Gaza between October 2023 and December 2025. Our method employs the CLIP ViT-B/32 visual encoder with selective fine-tuning of the final transformer block and a lightweight classification head. To address class imbalance, we apply focal loss along with standard data augmentation and threshold optimization. Experimental results show that the proposed model outperforms several CNN baselines and achieves an F1-score of 0.877 on the blind test set, securing 4th place in the shared task.

Stance detection has become an essential task for understanding political discourse on social media, particularly in highly polarized contexts where sentiment alone is insufficient to capture author intent. This study addresses stance classification in discussions related to the Palestinian–Israeli conflict by developing transformer-based and ensemble learning approaches for three-class classification: Pr-Palestine, Pro-Israel, and Neutral. Using the StanceNakba 2026 Shared Task dataset, we fine-tune multiple pretrained transformer models, including MARBERT, ARBERT, BERT, RoBERTa, and DeBERTa, and evaluate their performance using stratified cross-validation with macro F1-score as the primary metric. In addition to individual model evaluation, a weighted ensemble combining BERT, RoBERTa, and DeBERTa is proposed to leverage complementary contextual representations. Experimental results show that the ensemble model achieves the best performance with an accuracy and macro F1-score of 0.8905, outperforming specialized Arabic models while maintaining strong class-wise balance. The proposed approach achieved first place on the Codabench leaderboard in both the development and final evaluation phases of the shared task, demonstrating its robustness and effectiveness in real-world stance detection settings.

This research investigates multilingual propaganda detection by employing transformer-based models, specifically mBERT, XLM-RoBERTa, and mT5. The study utilizes a balanced dataset from the BiasFigNews corpus, annotated for propaganda and bias across five languages. The models were finely tuned to generate embeddings for classification tasks. The evaluation revealed mT5 as the most effective model, achieving an accuracy of 99.61% and an F1-score of 0.9961, followed by mBERT and XLM-RoBERTa with accuracies of 92% and 91.41%, respectively. The findings demonstrate the efficacy of transformer-based embeddings in detecting propaganda while also highlighting challenges in subtle class distinctions. Future work aims to enhance cross-lingual adaptability and explore lightweight models for resource-constrained settings.