Fable
The tax office finds the name that truly needs to be registered from a long letter
The county office needed to extract names, places, goods, amounts and dates from report papers. When the clerk read only the full text, he could understand the gist, but found it hard to search by object for clues about the same case across different documents.
The constable had the clerk circle each candidate entity and label its type, then link the occurrence back to the original text. A "Chang'an" might be a place name; a "Chang'an Hao" might be a business name. Type can't be decided by surface form alone.
The same business name has old forms and abbreviations, and people with the same name live on different streets. The workbench adds aliases, context, and unique identifiers, merging only when evidence is sufficient, and leaving uncertain cases in a pending-review state.
Listing entities alone isn't enough; the clerk records the relationships among people, goods, dates, and places as well. Each relationship keeps its source sentence and confidence score, so during an investigation you can go back to the original paper to verify.
Entity extraction turns unstructured material into searchable, linkable business objects, but the type system, normalization rules and human review are also part of system quality.