Named-entity recognition via Bumblebee.

Identifies named entities (people, organisations, locations, …) in natural-language text and returns each as a span with type and confidence. Backed by Bumblebee's Bumblebee.Text.token_classification/3 serving with span aggregation, loading a pre-trained transformer from Hugging Face.

Default model

The default is Davlan/bert-base-multilingual-cased-ner-hrl, trained on 10 high-resource languages (Arabic, German, English, Spanish, French, Italian, Latvian, Dutch, Portuguese, Chinese) with the standard four-class CoNLL-2003 tag set:

• :per — person
• :org — organisation
• :loc — location
• :misc — miscellaneous

Override with :model to use a language-specific or domain-specific NER model.

Optional dependency

Like Text.POS, NER requires the optional :bumblebee and :exla Hex packages. Without :bumblebee, extract/2 raises with a helpful message.

Cold start and caching

First call downloads the model (~700 MB for the multilingual default) and traces the inference graph. Subsequent calls hit a :persistent_term-cached serving. For production, start a named serving at boot:

serving = Bumblebee.Text.token_classification(model_info, tokenizer,
            aggregation: :same)
{:ok, _pid} = Nx.Serving.start_link(serving: serving, name: MyApp.NER)

Text.NER.extract(text, serving: MyApp.NER)

Result shape

%Text.NER.Entity{
  text: "Barack Obama",
  type: :per,
  start: 0,
  end: 12,
  score: 0.99
}

Spans are byte offsets into the original input.