Robust open intent classification in many-shot and few-shot scenarios.

Abstract

Open Intent Classification (OIC) focuses on accurately classifying known intents while identifying unknown intents, particularly in dialogue systems where novel intents frequently arise. Existing methods typically leverage large-scale datasets to fine-tune pre-trained BERT and derive representations for constructing decision boundaries. However, limited by the fixed feature dimensions, the model is underfitted in many-shot scenario and is overfitted in few-shot scenario, which poses a challenge to effectively model the data distribution and define precise decision boundaries. To tackle these challenges, we propose Robust Open Intent Classification (ROIC) for both settings. This involves two key components: distance-aware contrastive learning and decision boundary construction. To effectively fit the data distribution, we upgrade the pairwise distances to a complete distance matrix to capture global relationships, while also enhancing inter-class separability and intra-class aggregation through hard negative sampling mining. Furthermore, to construct precise decision boundaries, we propose feature space transformation to map representations into an appropriate space, ensuring robustness in both settings. Extensive experiments on standard benchmarks demonstrate the effectiveness of ROIC in both many-shot and few-shot scenarios.