Task-oriented dialogue systems are traditionally built on a closed-world assumption: they assume every user query can be mapped to a predefined set of known intents. When a user introduces a novel query or an irrelevant request, standard intent classifiers fail catastrophically. Instead of admitting ignorance, they force the out-of-scope (OOS) input into a known category with high confidence, leading to frustrating user experiences.
To build reliable conversational agents, models must learn to identify the boundaries of their own knowledge. This problem—known as out-of-scope or out-of-domain (OOD) detection—requires moving away from unconstrained classification toward architectures that can explicitly say "none of the above."
The Softmax Bottleneck
To understand why standard intent classifiers struggle with out-of-scope inputs, we have to look at how they partition the latent semantic space.
A typical deep learning classifier passes text embeddings through a linear layer followed by a softmax function. Geometrically, this divides the entire embedding space into $K$ infinite half-spaces, where $K$ is the number of known intents. Every single point in the vector space belongs to one of these $K$ regions. As noted by Congying Xia et al. and Ankan Mullick et al., this setup leaves no geometric "room" for unknown intents. If a user asks a chatbot trained only on weather and banking to "play music," the query's embedding will inevitably fall into either the weather or banking half-space, forcing an incorrect classification.
Bounding the Known Space
The most intuitive way to break the closed-world assumption is to replace infinite half-spaces with bounded decision regions. Instead of carving up the entire universe of text, the model draws a "safety bubble" around each known intent.
Standard Softmax (Infinite Regions) Bounded Boundaries (Safety Bubbles)
\ / --- ---
\ B / / B \ / C \
\ / \---/ \---/
A \ / C ---
------------*------------ / A \
/ \ \---/
/ \ [ Out-of-Scope Space ]
/ D \
One way to achieve this is through post-processing decision boundaries. Hanlei Zhang et al. developed the Adaptive Decision Boundary (ADB) method, which automatically learns a tight, closed boundary around each known intent cluster. If an incoming query falls outside all learned boundaries, it is automatically rejected as out-of-scope.
Another perspective leverages density estimation. Rather than drawing hard lines, models can learn the statistical distribution of known intents. In Revisiting Mahalanobis Distance for Transformer-Based Out-of-Domain Detection, the authors demonstrate that fine-tuning a Transformer encoder naturally groups known intents into compact clusters. By calculating the Mahalanobis distance from an input to these cluster centroids, the system can identify outliers based on how far they sit from the known distributions. Similarly, Guangfeng Yan et al. utilize a semantic-enhanced Gaussian Mixture Model (GMM) to map known intent densities, flagging queries that land in low-density regions as unknown.
Shaping the Latent Space via Contrastive Learning
Distance- and density-based detection methods rely on a critical assumption: that the underlying embedding space is well-segregated. If the representations of known intents are scattered or overlapping, out-of-scope queries cannot be cleanly separated.
To address this, researchers use contrastive learning to explicitly sculpt the latent space during training. Zhiyuan Zeng et al. employ Supervised Contrastive Learning (SCL) to pull embeddings of the same intent closer together while pushing different intents apart. This forces the model to compress known classes into hyper-dense, isolated islands, maximizing the empty space between them. When an out-of-scope query is introduced, it naturally lands in these vast unmapped gaps, making it geometrically obvious to distance metrics.
For scenarios where labeled data is scarce, Cheng Bing et al. apply an adaptive prototypical pseudo-labeling strategy (APP). This framework uses a push-pull mechanism on unlabeled data to maintain compact intent clusters, ensuring that even few-shot models retain clear boundaries between what they know and what they do not.
Manufacturing the Unknown
An alternative strategy treats out-of-scope detection as a standard classification task by introducing a $(K+1)$-th "unknown" class. The core challenge here is data scarcity: because the space of out-of-scope queries is infinite, it is impossible to collect a representative training set for the "unknown" category.
To circumvent this, models can synthesize proxy out-of-scope data. Zifeng Cheng et al. introduced a framework (SLMM) that uses Manifold Mixup to interpolate between different known intent representations. By blending these embeddings, the model generates synthetic points that lie in the ambiguous zones between known classes. Training the model to label these blended points as "unknown" teaches it to recognize the boundaries of its knowledge base without requiring real out-of-scope training examples.
Alternatively, the problem can be reframed from multi-class classification to pairwise matching. Jian-Guo Zhang 0005 et al. developed Discriminative Nearest Neighbor Classification (DNNC), which reformulates intent detection as a Natural Language Inference (NLI) task. Instead of asking "Which of these $K$ categories matches the input?", the model checks pairs of sentences to see if the user query entails any specific example in the training set. If the query fails to match any existing example with sufficient confidence, it is inherently classified as out-of-scope.
Managing Uncertainty and Clarification
In real-world deployments, the boundary between known and unknown is rarely a clean split. Queries are often ambiguous, underspecified, or partially out-of-scope. Hard rejections can frustrate users just as much as false acceptances.
To bridge this gap, systems can transition from single-guess predictions to set-valued predictions. Floris den Hengst et al. introduced Conformal Intent Classification and Clarification (CICC), which leverages conformal prediction to output a set of plausible intents that meet a strict statistical confidence guarantee. If the predicted set contains multiple intents, or if it returns empty, the system avoids making a blind guess. Instead, it triggers a targeted clarification question or a graceful fallback response, allowing the dialogue system to dynamically navigate the edges of its domain.
Key Foundational Work
To believe is to understand (1999) — contributed 1 concept elements to this direction.