Towards Trustworthy Dialogue Systems with Advanced Out-of-Scope Intent Detection Model

The growing demand for trustworthy dialogue systems emphasizes the need for consistently accurate responses to user inputs. The first step in developing a trustworthy dialogue system is detecting user inputs with the out-of-scope (OOS) intent. Advanced research on OOS intent detection enhances effectiveness by using data augmentation to generate numerous artificial OOS samples from a limited set of true OOS data, modelling its distribution for training. However, data augmentation presents challenges, including higher costs, increased time, and a greater risk of overfitting. Additionally, current studies treat the OOS intent as a homogeneous category equivalent to known intents within a classification framework, overlooking the inherent diversity of OOS intents. To tackle these challenges, we introduce a novel method called Anchor-Integrated Dynamic Out-of-scope Intent Learning (AIDOIL), which integrates the selected anchor to represent the OOS intent adapting to diverse inputs dynamically. The intent representations transform the global classification problem into a matching task that determines if a user input aligns with each intent. This eliminates the necessity to augment OOS data and accommodate the diversity of OOS intents through dynamic representation learning. We conducted extensive experiments on three public dialogue datasets, demonstrating that AIDOIL achieves an average 7.21% improvement in OOS detection accuracy, while maintaining an acceptable increase in training time.