In traditional machine learning, classification is often treated like a multiple-choice test where the options are meaningless numbers. If a model is tasked with identifying "Politics" vs. "Sports," it typically sees these as Label 0 and Label 1. To the model, these are abstract indices with no inherent relationship to the text they represent.
A growing body of research suggests this approach throws away the most valuable asset of modern NLP: the pre-trained semantic knowledge of language models. By replacing abstract indices with descriptive natural language, researchers are "grounding" labels in the same conceptual space as the input data.
The Index Problem vs. Semantic Anchoring
When we use arbitrary integers as labels, the model must learn the meaning of each category from scratch through statistical co-occurrence. This is particularly inefficient in low-resource settings. As Gaurav Singh and colleagues argue in Constructing Artificial Data for Fine-tuning for Low-Resource Biomedical Text Tagging with Applications in PICO Annotation, treating labels as natural language concepts allows a model to establish a direct representational link between the input and the target.
This "semantic anchoring" transforms the task. Instead of learning that "Diabetes" is Class 42, the model uses its pre-existing understanding of medical terms to recognize the category. Haoran Shi demonstrated this in Towards Automated ICD Coding Using Deep Learning, where encoding the literal textual definitions of medical codes allowed the system to bridge the gap between informal clinical shorthand and formal medical terminology.
Reframing Extraction as Conversation
One of the most powerful applications of this idea is reframing complex extraction tasks as Machine Reading Comprehension (MRC). In A Unified MRC Framework for Named Entity Recognition, Xiaoya Li and her team replaced traditional sequence labeling with natural language queries. Instead of tagging tokens, the model answers questions like "Which person is mentioned in the text?"
This works because the query itself contains the "annotation guidelines." By using the descriptive text of the guidelines as the prompt, the model leverages its internal knowledge of what a "person" or "organization" is, rather than relying on a rigid tagger. Similarly, Fayuan Li applied this to event extraction in Event Extraction as Multi-turn Question Answering, treating the identification of event arguments (who, where, when) as a dialogue.
Leveraging Pre-trained Semantic Priors
The intuition behind this shift is simple: pre-trained models like BERT or T5 already have a high-dimensional map of human language. When we use a label like PER, we are asking the model to learn a new, isolated point on that map. When we use the word "Person," we are pointing to a region the model has already explored.
Ben Athiwaratkun showed in Augmented Natural Language for Generative Sequence Labeling that replacing alphanumeric codes with descriptive terms significantly boosts performance in few-shot scenarios. The model can infer boundaries and categories based on the meaning of the label. This is echoed by Mo Yu in The Art of Prompting, where human-encoded prior knowledge in prompts proved more effective than learnable vector parameters.
Beyond Simple Classification
This paradigm extends beyond text-to-text tasks:
- Dialogue Systems: Hwaran Lee developed SUMBT, which encodes the literal text of slot names (e.g., "price range") to track user goals across different domains without retraining.
- Numerical Regression: In Contrastive Learning Drug Response Models from Natural Language Supervision, Kun Li 0009 treated numerical drug responses as natural language descriptions, using the high-dimensional space of language to organize continuous values more effectively.
- Acoustic Recognition: Yuan Xie used text templates to incorporate environmental metadata into underwater sound recognition in Underwater-Art, allowing the model to "understand" the physical context of a signal.
Why It Works: The Semantic Decision Boundary
Traditional classifiers draw boundaries between points in a feature space based on label IDs. Semantic grounding, however, allows the model to compare the context of the input directly against the definition of the label.
As Duo Chai explains in Description Based Text Classification with Reinforcement Learning, treating labels as instructions allows models to distinguish between entangled features—like different aspects of sentiment—by looking at the explicit descriptions of those features. By moving from "what is the label ID?" to "which description best matches this text?", we enable models to reason about their choices using the same linguistic tools they use to understand the world.