Prompt Based CVAE Data Augmentation for Few-Shot Intention Detection

Intent detection is an important task for AI assistants when communicating with users. However, in real life, the number of intents that need to be recognized in the intent recognition task continues to increase. It is often difficult to manually label new intents and new expressions over time, so newly added intents often only have a small number of manually labeled sentences, which is bad news for large intent detection models. To solve the few-shot intention detection challenge of new data, we propose a soft prompt based data augmentation model. We combine the Conditional Variational Auto Encoder(CVAE)model which can generate variational similar sentences, with the Prompt Tuning method, which is good at generating pseudo examples in few-shot conditions. We utilized the proposed generative model to generate pseudo-labeled data for few-shot intents to alleviate this problem. The proposed model can generate similar sentences for few-shot intention, thereby transforming the problem into traditional supervised learning. The problem is solved without changing the downstream model of the intent recognition task. The experimental study has shown that our method achieves promising results on public datasets and has practical significance.