In traditional machine learning, category labels are treated as meaningless, orthogonal tokens. Whether a class is named "Dog", "Class 1", or 0x7f, a standard cross-entropy classifier treats it identically: as an arbitrary index in a one-hot vector. The model must learn the entire conceptual meaning of the category purely from the statistical distributions present in the training data.
Modern architectures—ranging from dual-encoder vision-language models like CLIP to text-to-text language models—fundamentally change this paradigm. Instead of mapping inputs to abstract indices, they map inputs to natural language labels. By embedding the text of the label itself into a shared semantic space, the model can leverage its pre-trained world knowledge to understand what a category means before seeing a single training example.
This paradigm shift underpins "dataless" and zero-shot classification. For instance, foundational work by Ming-Wei Chang et al. in Importance of Semantic Representation: Dataless Classification and Yangqiu Song et al. in On Dataless Hierarchical Text Classification demonstrated that mapping documents and category names into a shared semantic space (such as Wikipedia concepts) allows a system to categorize text without any labeled training data.
The Core Vulnerability: Semantic Label Dependency
While this approach unlocks powerful zero-shot and few-shot capabilities, it introduces a critical vulnerability: semantic label dependency. Because the model relies on the linguistic meaning of the label to establish a classification boundary, its performance is strictly bounded by the descriptiveness and clarity of the label text.
If a dataset uses non-descriptive names, abstract codes, or cryptic identifiers, the semantic bridge collapses. As shown by Taihong Xiao et al. in Exploiting Category Names for Few-Shot Classification with Vision-Language Models, initializing classifier weights using the text embeddings of category names drastically improves few-shot performance. However, if the category names are arbitrary or non-descriptive, the model cannot leverage its prior knowledge, causing performance to revert to standard random initialization baselines.
This dependency means that the choice of surface-level text used to name a class is no longer a cosmetic detail—it is a core hyperparameter of the model's architecture.
How Label Sensitivity Manifests Across Paradigms
The impact of label descriptiveness propagates through several distinct machine learning workflows:
Retrieval-Based Classification
Some systems frame classification as a dense retrieval task, matching user queries directly to a library of label descriptions. In QAID: Question Answering Inspired Few-shot Intent Detection, Asaf Yehudai demonstrates that when intent labels are non-descriptive or numeric, the model loses the benefit of pre-trained semantic overlap, severely degrading few-shot performance. Similarly, the framework proposed in Dense Retrieval as Indirect Supervision for Large-space Decision Making is strictly bounded by the quality of the textual descriptions assigned to each category; ambiguous definitions directly impair the model's ability to map inputs to the correct targets.
Prompt Generation and Text-to-Text Models
In generative classification, where models output the literal text of a label, the linguistic properties of the label dictate success. Chen Qiu et al. observe in Federated Text-driven Prompt Generation for Vision-Language Models that text-conditioned prompt generation fails if class names lack descriptive detail. Furthermore, in ShefCDTeam at SemEval-2024 Task 4: A Text-to-Text Model for Multi-Label Classification, Meredith Gibbons notes that text-to-text models excel when labels correspond to explicit, easily identifiable textual features, but struggle significantly when forced to generate abstract rhetorical concepts that lack direct linguistic markers.
Out-of-Scope Detection
Label clarity also defines the boundaries of what a model does not know. Chandan Gautam et al. show in Class Name Guided Out-of-Scope Intent Classification that aligning queries with class names in a shared latent space creates tight boundaries around known intents. If the class names are non-descriptive, the model loses its semantic anchors, failing to distinguish between known and out-of-scope inputs.
Mitigating the Dependency with Richer Semantic Anchors
To overcome the limitations of fragile, single-word labels, recent research focuses on enriching the semantic context provided to the model. If a simple label name is insufficient, models can be guided by structural or definitional metadata.
One approach is to substitute or augment label names with full textual definitions. In Improving Few-Shot Relation Classification by Prototypical Representation Learning with Definition Text, Zhenzhen Li et al. train a prototype encoder directly on relation definitions rather than names alone, allowing the model to "read the manual" to form accurate prototypes even in niche domains.
Another strategy involves mapping the relationships between labels explicitly. Wenxuan Ma et al. introduce this in Language Semantic Graph Guided Data-Efficient Learning, constructing a Language Semantic Graph from natural language label descriptions. By using a Graph Neural Network to pass structural information between labels, they provide the primary model with a robust semantic framework. However, as the authors note, if the underlying labels are abstract or poorly defined, the resulting graph lacks the structural integrity needed to guide the model effectively.
Key Foundational Work
Importance of Semantic Representation: Dataless Classification (2008) — contributed 1 concept elements to this direction.
Zero-Shot Learning and Clustering for Semantic Utterance Classification (2014) — contributed 1 concept elements to this direction.
On Dataless Hierarchical Text Classification (2014) — contributed 1 concept elements to this direction.