Data annotation assigns human-provided labels or metadata to data examples. In supervised learning, these annotations become the targets a model learns to predict; when labels cannot be retrieved automatically, people must create them by hand. section 6.1.2 section 6.1.1
Annotation is often labor-heavy, and its quality directly affects target quality and therefore model quality. The workflow may use internal annotators, crowdsourcing, or a specialized labeling company, each trading control, cost, scale, and potential label noise. section 6.1.2
Match the annotation process to the task and its constraints:
- Use general annotators for straightforward categories such as cat-versus-dog image classification. section 6.1.2
- Use trained specialists when labels require domain knowledge, such as dog-breed classification or medical interpretation of CT scans. section 6.1.2
- Check whether labels can be generated automatically before committing to manual annotation. section 6.1.2
- Secure the required annotations before expecting a supervised model to learn the intended prediction task. section 6.1.1
From reviews to sentiment labels
A sentiment classifier needs both movie reviews and sentiment annotations. Without labeled examples, the intended prediction task lacks the training targets required for supervised learning. section 6.1.1
