Intent detection is a crucial task that bridges human intents and system actions in human-machine interaction systems. However, challenges remain in detecting out-of-scope (OOS) intents:
-
Traditional methods treat OOS intent detection as a multi-class classification problem, leading to decreased accuracy as the number of known intent classes increases;
-
LLM-embedding methods require large parameters, making them difficult to train and deploy in practice.
To address these issues, this work proposes a multi-cluster boundary learning method for OOS intent detection using MiniLM embedding (i.e., all-MiniLM-L6-v2) within a one-class classification framework. The method learns the boundaries of multi-cluster embeddings generated by MiniLM from training utterances and rejects out-of-domain utterances as OOS intents.
Experiments conducted on public datasets CLINC150, StackOverflow, and Banking77 demonstrate that this method achieves state-of-the-art performance in OOS intent detection compared to other baselines. Ablation studies further indicate that the employed MiniLM adapts well to the workflow and utterance embedding requirements. The code is available in the supplementary materials.
Blogger's Review: This study significantly enhances OOS intent detection performance by introducing a multi-cluster boundary learning approach, showing impressive adaptability and effectiveness, especially when dealing with a large number of intent categories. This method provides new insights for future intent detection technologies, warranting further exploration and application.