Ellipsoid-Based Decision Boundaries for Open Intent Classification

TL;DR

Open world textual intent classification demands robust recognition of known intents while rejecting unknowns. EliDecide presents ellipsoid boundaries per known class, parameterized by a learnable $n\times n$ matrix, and trained with a dual expansion-contraction loss to balance empirical and open-space risks; supervised contrastive learning enhances representation quality for boundary construction. The method achieves state-of-the-art results on Banking 2020 Efficient, OOS, StackOverflow, and a question classification dataset, with ablations confirming the geometric benefits of ellipsoids over balls and the effectiveness of pseudo-open samples. This work offers a practical, threshold-free boundary-based approach that generalizes to diverse open-world NLP tasks and reduces reliance on large language models for open-set detection.

Abstract

Textual open intent classification is crucial for real-world dialogue systems, enabling robust detection of unknown user intents without prior knowledge and contributing to the robustness of the system. While adaptive decision boundary methods have shown great potential by eliminating manual threshold tuning, existing approaches assume isotropic distributions of known classes, restricting boundaries to balls and overlooking distributional variance along different directions. To address this limitation, we propose EliDecide, a novel method that learns ellipsoid decision boundaries with varying scales along different feature directions. First, we employ supervised contrastive learning to obtain a discriminative feature space for known samples. Second, we apply learnable matrices to parameterize ellipsoids as the boundaries of each known class, offering greater flexibility than spherical boundaries defined solely by centers and radii. Third, we optimize the boundaries via a novelly designed dual loss function that balances empirical and open-space risks: expanding boundaries to cover known samples while contracting them against synthesized pseudo-open samples. Our method achieves state-of-the-art performance on multiple text intent benchmarks and further on a question classification dataset. The flexibility of the ellipsoids demonstrates superior open intent detection capability and strong potential for generalization to more text classification tasks in diverse complex open-world scenarios.

Figures (6)

• Figure 1: Open Intent Classification in Banking: Model classifies user requests into known intents (subsequently triggering corresponding services), or detects unknown (open) intents (handled as exceptions).
• Figure 2: An example of a known class's anisotropic feature distribution. The ball boundary excludes a significant portion of known samples as the cost of avoiding open ones inside, while the ellipsoid includes most of the known samples without many open ones inside due to geometric flexibility.
• Figure 3: The structure of our method. The first stage is to learn discriminative representations by SCL and then learn ellipsoid-based boundaries with expansion and contraction losses by using both positive and pseudo-negative samples.
• Figure 4: Coarse-grained confusion matrices with balls defined by different coverage fractions (CFs) and the ellipsoid (KCR=25%, seed=0).
• Figure 5: Comparison of F1 and Accuracy for different hyperparameters on OOS.