Efficient Multi-Task Learning for Social Media Stance Detection via Context-Aware LoRA Adaptation

Analyzing social media opinions requires efficient models capable of discerning nuanced opinions, including stance, sarcasm, and sentiment. This paper introduces a parameter-efficient multi-task learning framework leveraging Low-Rank Adaptation, known as LoRA, on a DeBERTa-v3 architecture. Our approach achieves robust performance across these related tasks while drastically reducing computational overhead by training only 7.3 million parameters, representing just 3.75% of the full model. The architecture uniquely integrates parent comment context, platform metadata embeddings, and dynamic data sampling. Evaluations across five standard benchmarks demonstrate strong results, achieving an accuracy of 81.76% for stance detection, 76.47% for sarcasm detection, and 95.18% for sentiment analysis. Notably, the framework achieves state-of-the-art performance on challenging political stance targets, including ’Hillary Clinton’ with an F1 score of 84.27% and ’Joe Biden’ with an F1 score of 84.90%, surpassing specialized models. With significantly reduced checkpoint sizes (371MB vs. 1.5GB) and practical training throughput (76.5 samples/sec), our work delivers a validated, efficient, and scalable solution essential for nuanced, large-scale social media opinion mining in real-world applications.