The Generative Shortfall: a comparative analysis of autoregressive vs. extractive architectures in ABSA
The results show that although generative SLMs can be efficiently deployed on resource-constrained edge devices, structurally regularized extractive encoders provide significantly better performance for precise opinion mining on highly imbalanced data.
Chathurvedi V. Rama Chandra, T. Padmavathy
· Frontiers in Artificial Inte... · 0 citations