Integrate Flair models

[julien-c]

either here or directly in our private repo api-inference

[stefan-it]

Really cool! I have actually worked on a small prototype that mimics the output of the current token classification models:

from flair.data import Sentence
from flair.models import SequenceTagger

tagger: SequenceTagger = SequenceTagger.load("dbmdz/flair-historic-ner-onb")

sentence: Sentence = Sentence(input_text)

# Also show scores for recognized NEs
tagger.predict(sentence, all_tag_prob=True, label_name="predicted")

response = []

entities = []

for span in sentence.get_spans("predicted"):
    idx = [token.idx for token in span.tokens]

    current_entity = {
        "entity_group": span.tag,
        "word": span.to_original_text(),
        "start": idx[0] - 1, # Because Flair starts at 1
        "end": idx[-1],
        "score": span.score
    }

    entities.append(current_entity)


response = {
    "entities": entities
}

return jsonify(response)

---

Bookmark for token classification API: https://api-inference.huggingface.co/docs/python/html/detailed_parameters.html#named-entity-recognition-ner-task

[julien-c]

Thanks! Yeah it looks pretty easy!

We'll discuss internally to figure out if we would implement this here or directly in our api-inference repo (which hosts the transformers model inference)

[julien-c]

was implemented in our api-inference repo (@Narsil @n1t0) so closing this here