Unknown Intention Detection Method Based on Dynamic Constraint Boundary

Open intent classification is a challenging task in dialogue systems. On the one hand, ensuring the classification quality of known intents is crucial. On the other hand, identifying open (unknown) intent during testing. Current models are limited in finding the appropriate decision boundary to balance the performances of both known and open intents. In this chapter, a post-processing method is proposed to learn the adaptive decision boundary (ADB) for open intent classification. The model is initially pre-trained using labeled known intent samples. Then, well-trained features to automatically learn the adaptive spherical decision boundaries for each known intent. Specifically, a new loss function is introduced to balance both the empirical risk and the open space risk. This method does not need open samples and is free from modifying the model architecture. This approach is surprisingly insensitive with less labeled data and fewer known intents. Extensive experiments conducted on three benchmark datasets demonstrate significant improvements compared to state-of-the-art methods.