FEAM: A Dynamic Prompting Framework for Sentiment Analysis with Hierarchical Convolutional Attention

Ziqian Lin¹Dongze Wu²Xiangbeng Yang²Lingying Li²Zhenkai Qin^2*

• ¹School of Public Administration, Guangxi Police College, nanning, China
• ²School of Information Technolog, Guangxi Police College, nanning, China

Sentiment analysis is a key task in natural language processing, supporting applications such as opinion mining, customer feedback interpretation, and social media monitoring. With the rise of pretrained language models, prompt-based tuning has become a popular paradigm due to its efficiency and adaptability. However, most existing approaches rely on static or handcrafted prompt templates, which often fail to generalize across diverse domains and nuanced sentiment expressions. To address this limitation, we propose FEAM (Fused Emotion-aware Attention Model), a dynamic prompting framework that selects input-specific soft prompts from a learnable prompt pool using a query-aware controller. These prompts are fused with contextual representations obtained via a pretrained BERT encoder, and further refined through a sentiment modulation layer, a multi-scale convolutional module, and a topic-aware attention mechanism. Experiments on four benchmark datasets show that FEAM achieves strong and consistent performance, reaching F1-scores of91.55% on Rest16, 92.83% on Laptop, 93.10% on Twitter, and 90.11% on FinancialPhraseBank. Extensive ablation and robustness studies further validate the contribution of each component and highlight FEAM's effectiveness in handling sentiment classification across varied and noisy textual domains.