Unsupervised category assignment
Classify text into predefined categories when labeled examples are unavailable.
AI & ML · Const
Zero-shot Learning
Zero-shot classification with embeddings assigns new text to the most similar candidate class without labeled training data. Embed each class name or short description, compare it with the input embedding, and choose the highest-similarity class.
How it works
Zero-shot classification uses an embedding model without labeled training examples. Represent each candidate class by its name or a short description, embed the new text and every class representation, then compare similarities and select the class with the highest score.
Syntax
client.embeddings.create(model="text-embedding-3-small", input=[...])Key points
- Embeddings support classification when no labeled examples are available.
- Create one embedding for each candidate class name or concise class description.
- Embed the incoming text and compare it against every candidate-class embedding.
- Assign the class whose embedding has the greatest similarity to the input embedding.
Common mistakes
- Zero-shot classification does not require labeled training data, but its quality depends on the class representations and similarity method.
Example: Embedding-based zero-shot classification
from openai import OpenAI
client = OpenAI()
response = client.embeddings.create(
model="text-embedding-3-small",
input=["cat", "dog", "vehicle"]
)
# Compare each input embedding with embeddings for class descriptions,
# then choose the class with the highest similarity.This example obtains embeddings that can be compared with embeddings for candidate class descriptions; classification selects the most similar class.
When to use it
Intent routing
Match incoming text to candidate intents represented by short descriptions.
Semantic classification
Select the closest semantic label from a fixed list of classes.
Related
Vector embeddingsembeddings APItext-embedding-3-smallzero-shot classification