Abstract: To address the issues of insufficient feature discriminability, lack of robustness, and semantic drift caused by large generative models in existing few-shot intent classification models, a few-shot intent classification model based on adversarially enhanced discriminative feature learning mechanism and hierarchical knowledge distillation was proposed. First, a gradient perturbation-based adversarial augmentation mechanism was introduced in the continuous feature space by constructing hard sample pairs, guiding the model to learn more discriminative and robust decision boundary features, thereby alleviating feature space distribution collapse and improving fine-grained semantic discrimination. Additionally, a hierarchical knowledge distillation strategy for multi-layer semantic alignment was designed, enabling knowledge transfer from the teacher model to the student model at both intermediate representation and output distribution levels, thereby improving the semantic expression stability and generalization capability of the model in low-resource scenarios. Comparative experiments were conducted on four public datasets BANKING77, MCID, HINT3, and HWU64 against 10 baseline models including Direct Fine-Tuning ++ (DFT++) and Dynamic Label name Refinement (DLR). The results demonstrate that the proposed model achieves higher accuracy and superior overall performance compared to the competing models.

Key words: few-shot intent classification, Fast Gradient Method (FGM), adversarial data augmentation, contrastive learning, discriminative feature learning mechanism, hierarchical knowledge distillation 

摘要： 针对现有少样本意图分类模型存在特征判别能力不足、鲁棒性缺失及生成式大模型易引发语义漂移问题，提出一种基于对抗增强的判别式特征学习机制和层级知识蒸馏的模型。首先在连续特征空间中引入基于梯度扰动的对抗增强机制，通过构造困难样本对，引导模型学习更具判别力与鲁棒性的决策边界特征，有效缓解特征空间中的分布坍塌问题，提升对细粒度语义的区分能力；同时，设计面向多层语义对齐的层级知识蒸馏策略，分别在中间表示与输出分布层面实现教师模型对学生模型的知识迁移，增强模型在低资源场景下的语义表达稳定性与泛化能力。在BANKING77、MCID、HINT3和HWU64 4个公开数据集上与DFT++（Direct Fine-Tuning ++）和DLR（Dynamic Label name Refinement）等10个基线模型进行了对比实验，结果表明所提模型的准确率均优于对比模型，综合性能更优。

关键词: 少样本意图分类, 快速梯度方法, 对抗数据增强, 对比学习, 判别式特征学习机制, 层级知识蒸馏