U4RASD at StanceNakba Shared Task: Data Augmentation and Auxiliary Objectives for Arabic Stance Detection

Abstract

This paper describes a submission to Track B of the StanceNakba Shared Task on Arabic cross-topic stance detection in the political domain. We investigate LLM-based data augmentation, auxiliary training objectives including contrastive and multi-task learning, zero-shot prompting, and a preliminary terminology-based clustering approach. Our final system, based on MARBERTv2 with dialect-aware LLM-based augmentation, achieved 86% macro-F1 on the blind test set and ranked 3rd out of 10 teams. Our results show that dialect-aware augmentation substantially improved performance in a low-resource Arabic stance detection setting, while not all auxiliary objectives or clustering-based strategies yielded consistent gains. We release our code at https://acr.ps/1L9B9Tw.

Resources

Details