Zero-Shot User Intent Detection via Augmented Conditional Variational Autoencoders

User intent detection plays a critical role in modern information retrieval and dialog question answering systems. It can provide intelligent retrieval results through understanding user query intent and serve as the semantic understanding in the human-machine conversation process. However, intent labeling is a time-consuming and labor-intensive work, and how to handle diverse and novel intentions is becoming an urgent and challenging task. In this paper, we propose a cvae-based architecture for the user intent detection which tries to make the latent vectors of training far away from the most similar misclassified intent labels in order to improve the detection accuracy. The cvae-based architecture includes two parts. Intent-AugCVAE extracts semantic features from utterance and is used to discriminate existing intents, while Intent-AugCVAE-ZSL transfers the learning ability of Intent-AugCVAE to discriminate emerging user intentions. Experimental results show that our method can not only discriminate existing intents, but also be able to discriminate emerging intents effectively when no labeled utterances exist.