A Multi-cluster Boundary Learning Method for Out-of-Scope Intent Detection via MiniLM Embedding

This paper frames out-of-scope (OOS) intent detection as a one-class classification problem. It uses the lightweight all-MiniLM-L6-v2 model to embed training utterances, learns boundaries around multiple embedding clusters, and rejects inputs outside the learned domain. Experiments are reported on CLINC150, StackOverflow, and Banking77, with the authors claiming state-of-the-art OOS detection performance against baselines. Ablation studies examine MiniLM’s suitability for the workflow and utterance-embedding requirements. The supplied abstract does not include the numerical results, baseline names, or detailed experimental settings.