오픈 의도 분류를 위한 탄력적인 타원형 경계 학습
📝 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.
💡 Analysis
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.
📄 Content
Open world classification is critical for robust real-world systems, as it enables accurate recognition of known classes while simultaneously rejecting anomalies to reduce openspace risks (Yang et al. 2024b). This capability is essential across diverse domains, including autonomous driving (Geiger, Lenz, and Urtasun 2012) and medical image analysis (Zimmerer et al. 2022). In the realm of natural language understanding, dialogue systems specifically require the ability to recognize the known user intents while detecting unknown ones, which is a significant capability for system robustness and continuous improvement (Lin and Xu Check payment status
Reset passcode
Make and send a physical card
The passcode doesn’t work.
I would like a physical card.
I want to eat something.
Model : Open Intent Classfication System : Action Implementation class. This formulation aligns with the established paradigm of Open Set Recognition (OSR) (Scheirer et al. 2012).
While the early open world classification methods (Bendale and Boult 2016; Hendrycks and Gimpel 2018;Liang, Li, and Srikant 2020) relied on class probability distributions, recent approaches utilize deep representations from pretrained models, which primarily fall into two categories. Scoring-based approaches (Zhou, Liu, and Qiu 2022;Zhou et al. 2023;Zhang et al. 2023b;Yang et al. 2024a;Gautam et al. 2024) construct discriminative feature spaces and identify unknown samples through statistic-based scoring functions like local density measures, but they require manual threshold tuning, limiting their practical use. In contrast, boundary-based methods (Zhang, Xu, and Lin 2021;Zhang et al. 2023a;Liu et al. 2023;Chen et al. 2024;Li et al. 2025) avoid thresholds by learning adaptive decision boundaries. However, current boundary-based approaches typically assume isotropic feature distributions, ignoring the directional variance inherent in real-world data. As illustrated in Fig- In this paper, we propose EliDecide, an ellipsoid-based decision boundary learning method for open-world classification. First, supervised contrastive learning is applied to a pre-trained model to obtain discriminative representations that serve as the basis for adaptive boundary construction. In the feature representation space, we construct ellipsoidshaped boundaries for each known class and regard the outside of all ellipsoids as the open space. Each ellipsoid is parameterized by a learnable nonsingular matrix encoding both the directions and the lengths of the ellipsoid’s axes. We optimize the boundaries using a dual loss mechanism. A positive loss expands the boundary when known samples (used as positive samples) fall outside it to encourage it to involve more known samples, while a negative loss contracts the boundary when pseudo-open samples (used as negative examples) lie too close or within it to prevent it from overexpanding and involving unknown samples. This integrated strategy enables EliDecide to adapt decision boundaries to the inherent directional variance in real-world data, representing the distributions of known samples flexibly and precisly and achieving robust open-world classification.
We extensively evaluate EliDecide on benchmark intent datasets and demonstrate its superior performance and robustness compared to state-of-the-art methods. Beyond intent, experiments on a question classification dataset show the generalization potential of our method in more diverse and complex open-world classification tasks. Additional experiments including comparisons among boundaries of different shapes further validate the advantage of the ellipsoid formulation, highlighting its effectiveness and robustness across different open-world scenarios.
We summarize our contributions as follows:
• To overcome the limitation of spherical boundaries in existing methods, we introduce the first ellipsoid boundary method for open world classification. We propose an effective parameterization that represents an ellipsoid via a learnable matrix capturing both axis directions and lengths. Novel and effective dual loss mechanism is designed to optimize the parameter matrices of boundaries.
Based on the primary detection criterion, existing methods fall into two categories: probability-based methods and deep representation-based methods.
Early-stage methods detect open samples using class probability distributions. For example, OpenMax (Bendale and Boult 2016) modifies softmax networks by replacing the penultimate layer with activation vectors to obtain the probabilities of (K + 1)-way classification task. Softmax-MSP (Hendrycks and Gimpel 2018) demonstrates that softmax probabilities can serve as a baseline for distinguishing between known and unknown data, while ODIN (Liang, Li, and Srikant 2020) improves this approach by applying temperature scaling and input perturbations. However, these methods rely solely on output probabilities and fail to leverage the deep semantic features.
Recent
This content is AI-processed based on ArXiv data.