Improving Optimal Prompt Learning through Multilayer Fusion and Latent Dirichlet Allocation

Qinghua Chen^1,2,3*Jessica Korneder⁴Osamah A Rawashdeh²Yanfeng Wang³Wing-Yue Geoffrey Louie⁵

• ¹Intelligent Robotics Laboratory, Oakland University, Rochester, United States
• ²Oakland University, Embedded Systems Research Lab, Rochester, United States
• ³Zhengzhou University of Light Industry, Zhengzhou, Henan Province, China
• ⁴Applied Behavior Analysis Clinic, Oakland University, Rochester, United States
• ⁵Oakland Universoty, Intelligent Robotics Laboratory, Rochester, United States

Recent advances in few-shot learning have demonstrated the potential of prompt-based techniques with pre-trained models, eliminating the need for extensive fine-tuning. However, challenges such as obtaining optimal prompts and addressing data scarcity in specialized domains remain challenging. To address these limitations, we propose a novel framework incorporating a Global Attention Mechanism (GAM) that effectively integrates features from multiple layers of pre-trained language models, augmented by Latent Dirichlet Allocation (LDA) generated topic features for prompt optimization. Extensive experiments on four datasets consistently show that our approach outperforms state-of-the-art baselines. The strategic integration of GAM with layer-specific features and LDA topics proves particularly effective in extracting valuable latent information for few-shot learning scenarios, yielding significant improvements in specialized domains, as evidenced by enhanced performance in therapeutic dialogue classification within a Applied Behavior Analysis clinical dataset.