In traditional machine learning, classification is treated as a mapping from an input to an arbitrary integer. To a model, "Label 0" and "Label 1" are meaningless indices; the model only learns what they represent by seeing thousands of labeled examples. This "one-hot" approach fails in the "zero-shot" regime, where a model must categorize data into classes it has never encountered during training.
A growing body of research suggests a more intuitive alternative: treating labels as semantic concepts rather than discrete IDs. By representing categories through natural language names, definitions, or even detailed narratives, researchers are enabling models to "read" a class description and immediately understand how to identify it.
The Shared Semantic Space
The core mechanism behind this shift is the projection of both inputs and labels into a shared embedding space. Instead of a fixed output layer, models use a dual-encoder or "Siamese" architecture to calculate the similarity between an instance and a candidate label.
Early work in this area, such as that by Ariadna Quattoni in Structured Prediction with Output Embeddings for Semantic Image Annotation, demonstrated that representing labels as feature vectors allows models to generalize to rare or unseen combinations of concepts. By anchoring these representations to pre-trained word embeddings, as explored by Erik Cambria and colleagues in Label Embedding for Zero-shot Fine-grained Named Entity Typing, a model can identify a "Corporation" even if it was only trained on the broader category "Organization," simply because the word vectors for those terms are semantically close.
From Labels to Natural Language Queries
While using a single word as a label is a start, it often lacks the nuance required for complex tasks like Event Extraction (EE). Researchers like Rui Feng and Chao Zhang have reframed these tasks as Machine Reading Comprehension (MRC). In Probing and Fine-tuning Reading Comprehension Models for Few-shot Event Extraction, they use the actual technical descriptions from annotation manuals as queries.
Similarly, Mo Yu and Lifu Huang proposed a "query-and-extract" framework in Query and Extract: Refining Event Extraction as Type-oriented Binary Decoding. By treating event types as natural language questions, the model leverages the vast linguistic knowledge it acquired during pre-training to "reason" about whether a sentence describes a specific event, even if that event type is entirely new.
Bridging Modalities with Textual Descriptions
This semantic bridging is not limited to text-to-text tasks. It has become a powerful tool for multi-modal learning, where text acts as the "universal translator" for other data types:
- Proteins: Minghao Xu and Jian Tang developed ProtST, which aligns protein sequences with biomedical text descriptions. This allows scientists to retrieve proteins using natural language prompts like "proteins involved in insulin resistance" without needing a specialized database of labels.
- Images: In ZEST, Tzuf Paz-Argaman and Yuval Atzmon use Wikipedia articles to classify bird species. By extracting visual "stories" from text, the model learns what a bird looks like (e.g., "red beak," "long tail") from a description before ever seeing a photo.
- Tabular Data: Alejandro Buendia and David Sontag showed in TabLLM that by serializing table rows into natural language sentences, Large Language Models can perform zero-shot classification by simply understanding the semantic meaning of the column headers.
Why It Works: The Power of Indirect Supervision
The effectiveness of this approach stems from the fact that modern models are pre-trained on massive corpora where labels and their meanings are naturally intertwined. As Xiaochuang Han and Yulia Tsvetkov discovered with ORCA, a model’s ability to perform a zero-shot task often depends on whether its pre-training data contained synonyms or descriptions of the target labels.
By using semantic representations, we stop asking the model to learn a new, isolated mapping for every task. Instead, we allow it to use its existing understanding of the world to navigate new domains. As Congying Xia and Wenpeng Yin argue in their work on Incremental Few-shot Text Classification, reframing classification as a semantic entailment problem—where the model decides if a text "matches" a label's definition—is the key to building AI that can learn and adapt as fluidly as humans do.