Definition
Zero-shot classification is a natural-language-processing technique that classifies text against labels the model has never seen training examples for. Instead of labeled data, the model consumes a natural-language description of each label and decides whether the description applies to the input. Modern zero-shot classifiers are built on either entailment models (e.g., BART-MNLI) or large language models prompted with the label set and a classification instruction.
The accuracy varies by label and by text domain. Well-described labels on clearly written text (a review that plainly says "the basket cracked") reach supervised-classifier accuracy. Ambiguous labels, niche jargon, or adversarial text (sarcasm, mixed aspects) degrade. The usual practice is to run zero-shot first, then collect labeled examples for the labels that matter most, then fine-tune or swap in a supervised classifier for those specific labels.
Why it matters
For consumer brands, feedback themes appear and disappear faster than any labeling program can keep up. A new defect on a launch SKU, a viral complaint about a packaging change, a regional issue with a specific retailer — each needs a label within days, not a quarter. Zero-shot classification is what makes that possible. A Consumer Insights analyst writes a one-sentence description of the new theme, the system starts classifying against it immediately, and the backfill over the prior 90 days of feedback runs in minutes.
The trade-off is evaluation discipline. A label that matters to the business deserves a labeled holdout set and accuracy measurement, even if it started zero-shot. Without that, nobody knows whether the label is working — only that records are being tagged. Indellia's Theme Agent pairs zero-shot theme creation with lightweight labeled-evaluation workflow, so the themes a brand relies on are the themes a brand has measured. Long-tail categories — niche defects, obscure accessories, uncommon customer segments — get coverage they would never justify in a supervised pipeline.
Example
A personal-care brand sees a novel theme emerging on Amazon reviews: "pump stops working after three weeks." The Consumer Insights analyst writes a one-sentence description and adds it as a zero-shot theme in Indellia. Within minutes, the Theme Agent classifies 1,240 reviews against the new label, including 180-day backfill. The analyst reviews a sample, marks 60 correct and 12 incorrect, and promotes the theme into the active taxonomy with a measured 83% precision. QA opens a ticket on the pump supplier that same afternoon. A traditional supervised pipeline would have needed a labeling project, a training run, and a deployment cycle — days or weeks during which the defect would have kept compounding.