README.md

metadata

license: apache-2.0
tags:
  - flair
  - token-classification
  - sequence-tagger-model
language: es
datasets:
  - conll2003
  - BSC-LT/NextProcurement-NER-Spanish-UTE-Company-annotated
widget:
  - text: >-
      PODACESA OBRAS Y SERVICIOS, S.A, y ECR INFRAESTRUCTURAS Y SERVICIOS
      HIDRÁULICOS S.L., constituidos en UTE PODACESA-ECR realizan la siguiente
      oferta:

Recognition of UTEs and company mentions in Flair

This is a model trained using Flair to recognise mentions of UTEs (Unión Temporal de Empresas) and companies in public tenders.

It is a finetune of the flair/ner-spanish-large model (retrained from scratch to include additional tags).

Results:
- F-score (micro) 0.7431
- F-score (macro) 0.7429
- Accuracy 0.5944

By class:
                precision    recall  f1-score   support

           UTE     0.7568    0.7887    0.7724        71
SINGLE_COMPANY     0.6538    0.7846    0.7133        65

     micro avg     0.7039    0.7868    0.7431       136
     macro avg     0.7053    0.7867    0.7429       136
  weighted avg     0.7076    0.7868    0.7442       136

Based on document-level XLM-R embeddings and FLERT.

---

Demo: How to use in Flair

Requires: Flair (pip install flair)

from flair.data import Sentence
from flair.models import SequenceTagger
# load tagger
tagger = SequenceTagger.load("BSC-LT/NextProcurement-NER-Spanish-UTE-Company")
# make example sentence
sentence = Sentence("PODACESA OBRAS Y SERVICIOS, S.A, y ECR INFRAESTRUCTURAS Y SERVICIOS HIDRÁULICOS S.L., constituidos en UTE PODACESA-ECR realizan la siguiente oferta:")
# predict NER tags
tagger.predict(sentence)
# print sentence
print(sentence)
# print predicted NER spans
print('The following NER tags are found:')
# iterate over entities and print
for entity in sentence.get_spans('ner'):
    print(entity)

This yields the following output (TODO: update):

Span [1,2]: "George Washington"   [− Labels: PER (1.0)]
Span [5]: "Washington"   [− Labels: LOC (1.0)]

So, the entities "George Washington" (labeled as a person) and "Washington" (labeled as a location) are found in the sentence "George Washington fue a Washington".

---

Training: Script to train this model

The following Flair script was used to train this model (TODO: update):

import torch
# 1. get the corpus
from flair.datasets import CONLL_03_SPANISH
corpus = CONLL_03_SPANISH()
# 2. what tag do we want to predict?
tag_type = 'ner'
# 3. make the tag dictionary from the corpus
tag_dictionary = corpus.make_tag_dictionary(tag_type=tag_type)
# 4. initialize fine-tuneable transformer embeddings WITH document context
from flair.embeddings import TransformerWordEmbeddings
embeddings = TransformerWordEmbeddings(
    model='xlm-roberta-large',
    layers="-1",
    subtoken_pooling="first",
    fine_tune=True,
    use_context=True,
)
# 5. initialize bare-bones sequence tagger (no CRF, no RNN, no reprojection)
from flair.models import SequenceTagger
tagger = SequenceTagger(
    hidden_size=256,
    embeddings=embeddings,
    tag_dictionary=tag_dictionary,
    tag_type='ner',
    use_crf=False,
    use_rnn=False,
    reproject_embeddings=False,
)
# 6. initialize trainer with AdamW optimizer
from flair.trainers import ModelTrainer
trainer = ModelTrainer(tagger, corpus, optimizer=torch.optim.AdamW)
# 7. run training with XLM parameters (20 epochs, small LR)
from torch.optim.lr_scheduler import OneCycleLR
trainer.train('resources/taggers/ner-spanish-large',
              learning_rate=5.0e-6,
              mini_batch_size=4,
              mini_batch_chunk_size=1,
              max_epochs=20,
              scheduler=OneCycleLR,
              embeddings_storage_mode='none',
              weight_decay=0.,
              )
)

---