Meta Learning for Few-Shot Joint Intent Detection and Slot-Filling

Abstract

Intent detection and slot filling are the two main tasks in natural language understanding module in goal oriented conversational agents. Models which optimize these two objectives simultaneously within a single network (joint models) have proven themselves to be superior to mono-objective networks. However, these data-intensive deep learning approaches have not been successful in catering the demand of the industry for adaptable, multilingual dialogue systems. To this end, we cast joint intent detection as an n-way k-shot classification problem and establish it within meta learning setup. Our approach is motivated by the success of meta learning on few-shot image classification tasks. We empirically demonstrate that, our approach can meta-learn a prior from similar tasks under highly resource constrained settings which enable rapid inference on target tasks. First, we show the adaptability of proposed approach by meta learning n-way k-shot joint intent detection using set of intents and evaluating on a completely new set of intents. Second, we exemplify the cross-lingual adaptability by learning a prior, utilizing English utterances and evaluating on Spanish and Thai utterances. Compared to random initialization, our method significantly improves the accuracy in both intent detection and slot-filling.