What it means
Classic classifiers need many labeled examples for each category. Zero-shot classifiers, and large language models used as classifiers, can sort items into new categories just from their names and descriptions, such as "billing question", "bug report" or "feature request".
Adding a few examples, called few-shot classification, often improves accuracy further.
Why it matters for a business
Zero-shot classification lets businesses sort messages, products or documents without building a training dataset first. Combined with simple rules for clear cases, it can sort large volumes cheaply and consistently.
A business example
Things to watch
-
Write clear, distinct category descriptions.
-
Measure accuracy on a labeled sample.
-
Use rules first for easy cases to save cost.
-
Route low-confidence items to review.