Logistic Regression-Based Example Selection for Enhanced Few-Shot Learning in Intent Classification

This study introduces a novel approach to example selection in few-shot learning scenarios for dialog intent classification, leveraging logistic regression to refine the set of examples retrieved through traditional similarity-based methods. We evaluate our method on three benchmark datasets: BANKING77, CLINC150, and HWU64, using 5-shot and 10-shot learning setups with 20 demonstrations. Our results show improvements in classification accuracy compared to baseline similarity-based retrieval methods, particularly for semantically similar intents. Notably, we observe a reduction in misclassifications within similar domains after applying our proposed approach. This work contributes to the growing body of research on efficient few-shot learning techniques for natural language understanding tasks, offering insights into enhancing performance in challenging, domain-specific scenarios.