SetFitQuad: A Few-Shot Framework for Aspect Sentiment Quad Prediction With Sampling Strategies

Aspect Sentiment Quad Prediction (ASQP) aims to extract structured sentiment quadruples, comprising aspect category, aspect term, opinion expression, and sentiment polarity from text to enable fine-grained sentiment analysis. Although generative models, such as T5, achieve strong results in this task, they require extensive labeled data and computational resources, limiting their practicality in low-resource scenarios. This study introduces SetFitQuad, a lightweight few-shot learning framework that decomposes ASQP into modular subtasks using sentence-embedding classification and syntactic-rule-based opinion extraction. To further enhance the performance on small and imbalanced datasets, SetFitQuad integrates a sampling-aware fine-tuning strategy based on entropy, density, and representational distance heuristics. Experiments on benchmark datasets show that SetFitQuad improves the F1-score for aspect category detection from 0.379 to 0.575 and sentiment polarity classification from 0.514 to 0.826, while reducing the training time by over 65% compared with the T5-based baseline. Furthermore, SetFitQuad achieves an F1-score of 0.532 for partial matches and 0.083 for exact matches on the ASQP task, demonstrating a competitive structured prediction performance under resource-conscious conditions.