Few-shot Learning for Multi-label Intent Detection

In this paper, we study the few-shot multi-label classification for user intent detection. For detection, state-of-the-art work estimates label-instance relevance scores and uses a threshold to select multiple associated labels. To determine appropriate thresholds with only few examples, first learn universal thresholding experience on data-rich domains, then adapt certain domains calibration based nonparametric learning. better calculation of score, introduce label name embedding as anchor points in representation space, which refines representations different classes be well-separated from each other. Experiments two datasets show that proposed model significantly outperforms strong baselines both one-shot five-shot settings.