mrm8488/electricidad-base-discriminator

ELECTRICIDAD: The Spanish Electra Imgur

Electricidad-base-discriminator (uncased) is a base Electra like model (discriminator in this case) trained on a Large Spanish Corpus (aka BETO's corpus)

As mentioned in the original paper: ELECTRA is a new method for self-supervised language representation learning. It can be used to pre-train transformer networks using relatively little compute. ELECTRA models are trained to distinguish "real" input tokens vs "fake" input tokens generated by another neural network, similar to the discriminator of a GAN. At small scale, ELECTRA achieves strong results even when trained on a single GPU. At large scale, ELECTRA achieves state-of-the-art results on the SQuAD 2.0 dataset.

For a detailed description and experimental results, please refer the paper ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators.

Model details ⚙

Name	# Value
Layers	12
Hidden	768
Params	110M

Evaluation metrics (for discriminator) 🧾

Metric	# Score
Accuracy	0.985
Precision	0.726
AUC	0.922

Fast example of usage 🚀

from transformers import ElectraForPreTraining, ElectraTokenizerFast
import torch

discriminator = ElectraForPreTraining.from_pretrained("mrm8488/electricidad-base-discriminator")
tokenizer = ElectraTokenizerFast.from_pretrained("mrm8488/electricidad-base-discriminator")

sentence = "El rápido zorro marrón salta sobre el perro perezoso"
fake_sentence = "El rápido zorro marrón amar sobre el perro perezoso"

fake_tokens = tokenizer.tokenize(fake_sentence)
fake_inputs = tokenizer.encode(fake_sentence, return_tensors="pt")
discriminator_outputs = discriminator(fake_inputs)
predictions = torch.round((torch.sign(discriminator_outputs[0]) + 1) / 2)

[print("%7s" % token, end="") for token in fake_tokens]

[print("%7s" % prediction, end="") for prediction in predictions.tolist()]

# Output:
'''
el rapido  zorro  marro    ##n   amar  sobre     el  perro   pere ##zoso    0.0    0.0    0.0    0.0    0.0    0.0    1.0    1.0    0.0    0.0    0.0    0.0    0.0[None, None, None, None, None, None, None, None, None, None, None, None, None
'''

As you can see there are 1s in the places where the model detected a fake token. So, it works! 🎉

Some models fine-tuned on a downstream task 🛠️

Question Answering

POS

NER

Spanish LM model comparison 📊

Dataset	Metric	RoBERTa-b	RoBERTa-l	BETO	mBERT	BERTIN	Electricidad-b
UD-POS	F1	0.9907	0.9901	0.9900	0.9886	0.9904	0.9818
Conll-NER	F1	0.8851	0.8772	0.8759	0.8691	0.8627	0.7954
Capitel-POS	F1	0.9846	0.9851	0.9836	0.9839	0.9826	0.9816
Capitel-NER	F1	0.8959	0.8998	0.8771	0.8810	0.8741	0.8035
STS	Combined	0.8423	0.8420	0.8216	0.8249	0.7822	0.8065
MLDoc	Accuracy	0.9595	0.9600	0.9650	0.9560	0.9673	0.9490
PAWS-X	F1	0.9035	0.9000	0.8915	0.9020	0.8820	0.9045
XNLI	Accuracy	0.8016	0.7958	0.8130	0.7876	0.7864	0.7878

Acknowledgments

I thank 🤗/transformers team for allowing me to train the model (specially to Julien Chaumond).

Citation

If you want to cite this model you can use this:

@misc{mromero2020electricidad-base-discriminator,
  title={Spanish Electra by Manuel Romero},
  author={Romero, Manuel},
  publisher={Hugging Face},
  journal={Hugging Face Hub},
  howpublished={\url{https://huggingface.co./mrm8488/electricidad-base-discriminator/}},
  year={2020}
}

Created by Manuel Romero/@mrm8488

Made with ♥ in Spain