Zero-Shot Classification with Embeddings

Overview

Zero-shot classification uses embeddings to categorize items without category-specific training examples. Works by computing similarity between item and category descriptions.

How It Works

1. Embed category descriptions
2. Embed item to classify
3. Find most similar category
4. Return classification

Example

categories = [
    "product complaint",
    "shipping inquiry",
    "billing question"
]

# Embed categories
cat_embeddings = [model.encode(c) for c in categories]

# Classify new message
message = "Where is my package?"
msg_embedding = model.encode(message)

# Find closest category
similarities = [cosine_similarity(msg_embedding, c) for c in cat_embeddings]
category = categories[np.argmax(similarities)]  # "shipping inquiry"

Advantages

• No training data needed
• Instant deployment
• Easy to add new categories
• Works across domains

Use Cases

• Customer support routing
• Content moderation
• Document classification
• Intent detection
• Product categorization

Pricing

Depends on embedding model used.