NER-fine-tuned-BETO: model fine-tuned from BETO for NER task.

Language: es Datasets:

• conll2002
• Babelscape/wikineural

Introduction

[NER-fine-tuned-BETO] is a NER model that was fine-tuned from BETO on the 2002 Conll and the WikiNEuRal spanish datasets. Model was trained on the Conll 2002 train dataset (~8320 sentences) and a bootstrapped dataset of WikiNEuRal, where we re-evaluate the dataset and only keep the sentences where all the labels matched the predictions made. Model was evaluated on the test dataset of Conll2002.

Training data

Training data was classified as follow:

Alongside the IOB formatting, this is:

• B-LABEL if the word is at the beggining of the entity.
• I-LABEL if the word is part of the entity name, but not the first word.

How to use NER-fine-tuned-BETO with HuggingFace

Load the model and its tokenizer :

from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer = AutoTokenizer.from_pretrained("NazaGara/NER-fine-tuned-BETO", use_auth_token=True)
model = AutoModelForTokenClassification.from_pretrained("NazaGara/NER-fine-tuned-BETO", use_auth_token=True)

nlp = pipeline('ner', model=model, tokenizer=tokenizer, aggregation_strategy="simple")
nlp('Ignacio se fue de viaje por Buenos aires')

[{'entity_group': 'PER',
  'score': 0.9997764,
  'word': 'Ignacio',
  'start': 0,
  'end': 7},
 {'entity_group': 'LOC',
  'score': 0.9997932,
  'word': 'Buenos aires',
  'start': 28,
  'end': 40}]

Model Performance

Overall

By classes