Discovering new intents via spatio-temporal pseudo-label denoising

New intent discovery, which aims to identify unknown intents from unlabeled data, is crucial for information processing. Existing methods typically adopt a semi-supervised paradigm by leveraging pseudo-labels to enhance intent recognition. However, pseudo-labels are prone to noise, which can degrade model convergence and compromise recognition accuracy. To address this issue, we propose a novel framework that dynamically refines pseudo-labels by incorporating spatio-temporal features. Specifically, from a spatial perspective, we evaluate sample-wise confidence and inter-sample cohesion to assess pseudo-label reliability. From a temporal perspective, we track category consistency and distribution stability across sample groups to adapt to evolving data patterns. By integrating these features with an adaptive thresholding strategy, our framework effectively filters and corrects erroneous pseudo-labels. Experiments on five diverse benchmarks demonstrate that our method achieves state-of-the-art performance, providing a more robust solution for new intent discovery.