Zero-Shot Classification

Zero-shot learning refers to a model making predictions at inference time without relying on any labeled samples for the task, directly based on pre-trained knowledge or external structural information.

Zero-shot classification relies on the semantic understanding capabilities of pretrained large language models such as NLI variants of BART and RoBERTa.

The study proposes a training-free six-class wafer map defect classification framework that leverages a frozen large multimodal model and achieves zero-shot classification through connected-component highlighting and rule-ordering prompts.

The framework requires no labeled data or fine-tuning, because the model parameters are completely frozen, and only the prompt rules need to be adjusted to adapt to defect pattern drift caused by changes in production conditions.

The method has so far been validated only on two public datasets, and complex noise and rare defect patterns in actual production lines may pose challenges; rule order tuning still requires domain knowledge, and automated search for the optimal rule order is a future direction.

The paper is currently validated only on retrieval tasks and does not yet cover generation tasks or zero-shot classification.

LLM content moderation is more sensitive to explicit profanity, hate speech, and other semantic violations, but has insufficient ability to identify violation types that require external fact-checking, such as disguised links and doxxing information.

The weakness of zero-shot classification is its instability in judging categories with ambiguous boundaries or domain-specific terminology, with inference costs growing linearly with traffic and poor result interpretability and reproducibility.

Zero-shot models require no labeled data; simply writing out labels in natural language is enough to produce classification results.

The quality ceiling of weak supervision labels depends on the accuracy of the zero-shot model itself; if the model makes systematic errors, the training set will inherit that bias.

The recommended path is to use zero-shot for rapid prototyping in the early stage, use zero-shot outputs as weak annotations to build an initial training set in the middle stage, and migrate to a dedicated classifier in the mature stage to reduce costs and improve efficiency.

When data is sufficient, categories are stable, cost sensitivity is high, and traffic is large, training a classifier is preferred; when there is no data, rapid validation is needed, or categories change frequently, zero-shot is a more appropriate starting point.

Zero-shot classifiers (such as solutions based on CLIP or NLI) allow users to describe categories in natural language, offering greater flexibility but typically requiring a larger base model.

In zero-shot classification, the scores of all candidates sum to 1; the fewer the candidates, the larger the probability space available to each, so the threshold parameter needs to be calibrated individually based on the specific number of categories.

Laya adopts a zero-shot classification paradigm, whose underlying mechanism is typically based on a pre-trained language model fine-tuned on Natural Language Inference (NLI) tasks.

The study uses unsupervised topic modeling to automatically identify discussion topics from review text, then applies zero-shot classification to assign topics to the five food access dimensions.

The machine classification results achieved an agreement rate of 85.4% compared with manual coding.