A2NLP at StanceNakba Shared Task: Fine-Tuned AraBERT for Topic-Based Arabic Stance Detection

Abstract

AbstractThis paper describes A2NLP’s system for Subtask B of the StanceNakba Shared Task, which addresses cross-topic Arabic stance detection. The goal is to classify sentence–topic pairs into pro, against, or neutral labels. We introduce a topic-conditioned prompting strategy built on AraBERTv0.2-Twitter, where each instance is reformulated into a structured prompt that explicitly models the interaction between the sentence and its target topic. The model is trained using 5-fold stratified cross-validation with class-weighted loss to ensure robustness under mild label imbalance. Our final submission achieves a Macro-F1 score of 0.8483 on the official test set, outperforming the AraBERTv2 baseline (0.810) and ranking fifth overall. Ablation analysis confirms that topic-conditioned prompting substantially improves generalization across topics. The findings demonstrate the importance of structured input design and domain-aligned pretraining for reliable stance detection in dialectal Arabic social media discourse.

Resources

Details