Modular Multi-Task Learning for Emotion-Aware Stance Inference in Online Discourse

Stance detection on social media is increasingly vital for understanding public opinion, mitigating misinformation, and enhancing digital trust. This study proposes a modular Multi-Task Learning (MTL) framework that jointly models stance detection and sentiment analysis to address the emotional complexity of user-generated content. The architecture integrates a RoBERTa-based shared encoder with BiCARU layers to capture both contextual semantics and sequential dependencies. Stance classification is reformulated into three parallel binary subtasks, while sentiment analysis serves as an auxiliary signal to enrich stance representations. Attention mechanisms and contrastive learning are incorporated to improve interpretability and robustness. Evaluated on the NLPCC2016 Weibo dataset, the proposed model achieves an average F1-score of 0.7886, confirming its competitive performance in emotionally nuanced classification tasks. This approach highlights the value of emotional cues in stance inference and offers a scalable, interpretable solution for secure opinion mining in dynamic online environments.