Entity annotation is the marking up of entities inside raw text: people, organisations, places, dates, products, monetary amounts. What gets recorded is where the mention starts and ends plus which type it belongs to. The result is a structured layer a model or a search system can use directly.
Annotation follows a label scheme. Standard types such as person, organisation and location sit alongside project-specific ones. An ecommerce project adds brand, model code and size, a legal project adds clause number and court name. The same text is often labelled by several people and their agreement is measured, because an unclear scheme produces different decisions from different annotators.
This data becomes the training set for named entity recognition models. The model learns from thousands of hand-marked examples and then finds the same entities in text it has never seen.
A concrete case: an insurer's claim reports mention vehicle make, plate, incident date and location inside free-form text. A few thousand reports are annotated by hand, the trained model then extracts those fields from incoming reports, and opening a file stops being a manual data entry job.
Annotation quality shows up directly in model quality. On a set where boundaries were drawn inconsistently, the model learns inconsistent boundaries.