⚡️ Saturday AI Sparks 🤖 - 🏷️ Zero-Shot Classification (Classify without training)

Normally, text classification requires a labeled dataset and a training process. But what if you want to classify text into categories without training at all?

That’s where Zero-Shot Classification comes in. Using Hugging Face’s facebook/bart-large-mnli model, you can assign custom labels to text on the fly.

Why Zero-Shot Classification?

• No training needed → You just provide text + candidate labels.
• Flexible → Works with any label set you define (e.g., “positive”, “negative”, “neutral” or “acting”, “cinematography”, “soundtrack”).
• Fast prototyping → Great for tagging, routing, or feedback analysis.

Installing Requirements

pip install transformers torch

Minimal Implementation

Load the pipeline and run classification:

from transformers import pipeline

classifier = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")

text = "I loved the cinematography, but the plot was too slow and predictable."
candidate_labels = ["positive", "negative", "story", "acting", "cinematography", "soundtrack"]

result = classifier(text, candidate_labels)

This returns a dictionary with labels and scores ranked by confidence.

Multi-Label Option

If a text can belong to more than one category, enable multi-label:

result = classifier(text, candidate_labels, multi_label=True)

This gives independent confidence scores for each label.

Batch Classification

You can also pass a list of texts at once. The pipeline returns a list of results, one per input.

Sample Output

Single-label result (top label): cinematography | score: 0.8421

All labels ranked:
cinematography  -> 0.8421
negative        -> 0.6910
story           -> 0.6395
positive        -> 0.3212
acting          -> 0.1453
soundtrack      -> 0.1024

Key Takeaways

• Zero-shot classification lets you skip training while still categorizing text.
• Candidate labels are defined on the fly, making it highly flexible.
• Great for early experiments, tagging pipelines, and quick insights before building supervised models.