Towards Robust Arabic Stance Detection: Cross-Domain Generalization and Joint Training for Misinformation Mitigation

Abstract

Stance detection is a pivotal component of automated fact-checking and misinformation mitigation. Machine learning research on Arabic stance detection largely focuses on in-domain evaluations, which leaves a significant gap in understanding how models generalize across the diverse linguistic patterns found in the real-world news ecosystem. In this paper, we present a systematic evaluation of ten Arabic Transformer models across two publicly available benchmarks: AraStance and UnifiedFC. We investigate cross-domain transferability and the impact of training distributions on model robustness. Our results reveal a significant performance degradation during domain shifts and uncover a generalization disparity, where training on a more diverse dataset can outperform in-domain training for minority classes in specialized domains. Furthermore, we demonstrate that joint training can improve model resilience without incurring significant performance costs on the source domain. Finally, we identify the Discuss class as a bottleneck in Arabic stance detection. This work offers a leaderboard for Arabic stance detection and provides insights for the development of robust, universal misinformation detection systems.