Graphire: Novel Intent Discovery with Pretraining on Prior Knowledge using Contrastive Learning

In this paper, we introduce Graphire, an intent discovery system leveraging pretraining on predefined intents to automatically discover novel intents for intelligent personal assistants (IPA). In order to transfer the prior knowledge of predefined intents, Graphire first transforms predefined class memberships into pairwise relationships, and then learns a Siamese Neural Network (SNN) model classifying if two utterances have the same intent. The Siamese neural network condenses the prior knowledge of predefined intents in the form of trained neural-network weights and infers pairwise relationships among new utterance pairs. The contribution of the paper is threefold: (1) it proposes a pretraining paradigm based on contrastive learning to distill prior knowledge from existing intents; (2) it proposes a new method to discover novel intents, leveraging the prior knowledge; (3) it proposes a cluster summarization approach to assign labels to the intents. The experimental results demonstrate the effectiveness of pretraining in Graphire and Graphire’s capability to discover novel intents on a real-world IPA dataset with intents from disparate domains.