Named entity recognition (NER)
DefinitionNamed entity recognition finds mentions of real-world things in text, such as people, places and organisations, and labels each with its type. Training data for it is made by marking spans in sentences.
Making NER data
- Choose a small set of entity types.
- Write guidelines that say where a span starts and ends, and what to do with titles, nested names and ambiguous cases.
- Have several annotators mark the same texts.
- Keep spans a majority agreed on, and review the disagreements.
Where agreement breaks down
Most disagreement is about boundaries, not types: whether Dr. is part of the name, whether the university is part of the place. Worked examples in the guidelines remove most of it.
How Azmaval uses it
The text span question in Azmaval lets reviewers select text and pick a label, with optional relations between spans. Exports include character offsets, Label Studio JSON and CoNLL.
Frequently asked questions
How many labelled sentences does a NER model need?
It depends on the language, the model and the label set. A few thousand sentences is a common start for fine-tuning a pretrained model, and low-resource languages often need more care in annotation than more volume.
Can reviewers pre-label with a model?
Yes. A hidden column with predicted spans can pre-fill suggestions that reviewers correct.