Boosting Low-Resource Intent Detection with in-Scope Prototypical Networks

Identifying intentions from users can help improve the response quality of task-oriented dialogue systems. How to use only limited labeled in-domain (ID) examples for zero-shot unknown intent detection and few-shot ID classification is a more challenging task in spoken language understanding. Existing related methods heavily rely upon the multi-domain datasets containing large-scale independent source domains for meta-training. In this paper, we propose a universal In-scope Prototypical Networks for low-resource intent detection to be general to dialogue meta-train datasets lacking widely-varying domains, which focuses on the scope of episodic intent classes to construct meta-task dynamically. Also, we introduce loss with margin principle to better distinguish samples. Experiments on two benchmark datasets show that our model consistently outperforms other baselines on zero-shot unknown intent detection without deteriorating the competitive performance on few-shot ID classification.