Structural knowledge-guided complex-noise handling for open intent classification

Open intent classification aims to assign known intents to their corresponding classes while identifying unknown intents. Its success largely relies on accurately annotated data. However, real-world data often contains both in-distribution (IND) noise and out-of-distribution (OOD) noise, which degrades model performance. Despite recent progress, open intent classification under such complex-noise scenarios remains underexplored. Moreover, existing noise-handling methods face several limitations: 1) they typically rely on limited class-level information, such as prototypes, for noise detection, overlooking intra-class structural characteristics; 2) they commonly adopt global criteria for sample selection; 3) they tend to ignore the potential value of OOD noise in training. These limitations reduce noise detection accuracy and hinder representation learning. To address these challenges, we propose a structural knowledge-guided Complex-Noise Handling method for Open Intent Classification (CNOIC). Specifically, we introduce structural knowledge representation to capture the structural characteristics of each class via granular-balls. Furthermore, we design structure-guided criteria, derived from granular-ball attributes, to distinguish clean samples, IND noise, and OOD noise. Additionally, we present an open space-aware representation learning strategy that uses OOD noise prototypes as placeholders for unknowns, effectively expanding the unknown space. Experiments on three datasets validate the effectiveness of our proposed approach. The source code of CNOIC is available at: https://github.com/Liyanhuaa/CNOIC.