Few-Shot Intent Detection by Data Augmentation and Class Knowledge Transfer

Task-oriented dialogue systems find extensive applications in domains such as customer service, virtual assistants, and interactive information retrieval, in which intent detection stands as a key component for understanding user objectives and facilitating efficient interaction. However, practical applications often pose challenges to intent detection due to data scarcity and imbalance. To address these challenges, this paper proposes a few-shot intent detection approach, FSDA-CKT(Few-Shot Intent Detection by Data Augmentation and Class Knowledge Transfer), which enhances data through constructing sentence pairs and utilizes a class knowledge transfer structure to transfer class knowledge between head and tail classes. This method effectively tackles the issue of imbalanced class data in training samples and ensures high intent detection performance in few-shot scenarios. Extensive experimental results on four baselines show that, FSDA-CKT improves the accuracy of indomain(IND) intent detection while maintaining relatively high out-of-domain(OOD) intent detection performance.