Variable-Proxy Multi-Prototype RMD for Out-of-Distribution Intent Detection

​

Intent classification is an essential component of task-oriented dialogue systems. However, in real-world applications, user queries may include unknown intents that are not defined during training. Detecting such out-of-distribution (OOD) intents is particularly difficult when unknown intents are semantically close to known intents, as in fine-grained single-domain datasets such as BANKING77. In this study, I propose a Variable-Proxy Multi-Prototype Relative Mahalanobis Distance (VP-MP-RMD) method for OOD intent detection. The proposed method first divides samples within each known class into proxy groups based on BERT feature representations. Then, BERT is trained using Cross Entropy and Variable-Proxy Supervised Contrastive Learning, where only samples belonging to the same class and the same proxy are treated as positive pairs. This preserves intra-class semantic substructures. After training, adaptive multi-prototypes are reconstructed from the learned feature space, and OOD scores are computed using Relative Mahalanobis Distance, which subtracts the background distribution distance from the nearest prototype distance. Experiments on BANKING77 show that the proposed method improves near-OOD detection compared with MSP, Energy, Mahalanobis distance, and standard supervised contrastive learning baselines.