A Triplet-Contrastive Representation Learning Strategy for Open Intent Detection

Open intent detection aims to correctly classify known intents and identify unknown intents that never appear in training samples, thus it is of practical importance in dialogue systems. Discriminative intent representation learning is a key challenge of open intent detection. Previous methods usually restrict known intent features to compact regions to learn the representations, which assumes that open intent is outside regions. However, open intent can be distributed among known intents. To address this issue, this paper proposes a triplet-contrastive learning strategy to learn discriminative semantic representations and differentiate between similar open intents and known intents. Further, a method named Triplet-Contrastive Adaptive Boundary (TCAB) is proposed, which leverages the triplet-contrastive learning strategy and an adaptive decision boundary method to detect open intent. Extensive experiments on three benchmark datasets show that our method achieves substantial improvements compared with a list of baseline methods.