PlanTL-GOB-ES / roberta-large-bne

RoBERTa large trained with data from the National Library of Spain (BNE)

Table of Contents

Overview

• Architecture: roberta-large
• Language: Spanish
• Task: fill-mask
• Data: BNE

Model description

The roberta-large-bne is a transformer-based masked language model for the Spanish language. It is based on the RoBERTa large model and has been pre-trained using the largest Spanish corpus known to date, with a total of 570GB of clean and deduplicated text processed for this work, compiled from the web crawlings performed by the National Library of Spain (Biblioteca Nacional de España) from 2009 to 2019.

Intended uses and limitations

The roberta-large-bne model is ready-to-use only for masked language modeling to perform the Fill Mask task (try the inference API or read the next section). However, it is intended to be fine-tuned on non-generative downstream tasks such as Question Answering, Text Classification, or Named Entity Recognition. You can use the raw model for fill mask or fine-tune it to a downstream task.

How to use

Here is how to use this model:

>>> from transformers import pipeline
>>> from pprint import pprint
>>> unmasker = pipeline('fill-mask', model='PlanTL-GOB-ES/roberta-large-bne')
>>> pprint(unmasker("Gracias a los datos de la BNE se ha podido <mask> este modelo del lenguaje."))
[{'score': 0.0664491355419159,
  'sequence': ' Gracias a los datos de la BNE se ha podido conocer este modelo del lenguaje.',
  'token': 1910,
  'token_str': ' conocer'},
 {'score': 0.0492338091135025,
  'sequence': ' Gracias a los datos de la BNE se ha podido realizar este modelo del lenguaje.',
  'token': 2178,
  'token_str': ' realizar'},
 {'score': 0.03890657424926758,
  'sequence': ' Gracias a los datos de la BNE se ha podido reconstruir este modelo del lenguaje.',
  'token': 23368,
  'token_str': ' reconstruir'},
 {'score': 0.03662774711847305,
  'sequence': ' Gracias a los datos de la BNE se ha podido desarrollar este modelo del lenguaje.',
  'token': 3815,
  'token_str': ' desarrollar'},
 {'score': 0.030557377263903618,
  'sequence': ' Gracias a los datos de la BNE se ha podido estudiar este modelo del lenguaje.',
  'token': 6361,
  'token_str': ' estudiar'}]

Here is how to use this model to get the features of a given text in PyTorch:

>>> from transformers import RobertaTokenizer, RobertaModel
>>> tokenizer = RobertaTokenizer.from_pretrained('PlanTL-GOB-ES/roberta-large-bne')
>>> model = RobertaModel.from_pretrained('PlanTL-GOB-ES/roberta-large-bne')
>>> text = "Gracias a los datos de la BNE se ha podido desarrollar este modelo del lenguaje."
>>> encoded_input = tokenizer(text, return_tensors='pt')
>>> output = model(**encoded_input)
>>> print(output.last_hidden_state.shape)
torch.Size([1, 19, 1024])

Limitations and bias

At the time of submission, no measures have been taken to estimate the bias and toxicity embedded in the model. However, we are well aware that our models may be biased since the corpora have been collected using crawling techniques on multiple web sources. We intend to conduct research in these areas in the future, and if completed, this model card will be updated.

Training

Training data

The National Library of Spain (Biblioteca Nacional de España) crawls all .es domains once a year. The training corpus consists of 59TB of WARC files from these crawls, carried out from 2009 to 2019.

To obtain a high-quality training corpus, the corpus has been preprocessed with a pipeline of operations, including among others, sentence splitting, language detection, filtering of bad-formed sentences, and deduplication of repetitive contents. During the process, document boundaries are kept. This resulted in 2TB of Spanish clean corpus. Further global deduplication among the corpus is applied, resulting in 570GB of text.

Some of the statistics of the corpus:

Training procedure

The training corpus has been tokenized using a byte version of Byte-Pair Encoding (BPE) used in the original RoBERTA model with a vocabulary size of 50,262 tokens.

The roberta-large-bne pre-training consists of a masked language model training, that follows the approach employed for the RoBERTa large. The training lasted a total of 96 hours with 32 computing nodes each one with 4 NVIDIA V100 GPUs of 16GB VRAM.

Evaluation

When fine-tuned on downstream tasks, this model achieves the following results:

For more evaluation details visit our GitHub repository or paper .

Additional information

Author

Text Mining Unit (TeMU) at the Barcelona Supercomputing Center ( [email protected] )

Contact information

For further information, send an email to [email protected]

Copyright

Copyright by the Spanish State Secretariat for Digitalization and Artificial Intelligence (SEDIA) (2022)

Licensing information

This work is licensed under a Apache License, Version 2.0

Funding

This work was funded by the Spanish State Secretariat for Digitalization and Artificial Intelligence (SEDIA) within the framework of the Plan-TL.

Citation information

If you use this model, please cite our paper :

@article{,
   abstract = {We want to thank the National Library of Spain for such a large effort on the data gathering and the Future of Computing Center, a
Barcelona Supercomputing Center and IBM initiative (2020). This work was funded by the Spanish State Secretariat for Digitalization and Artificial
Intelligence (SEDIA) within the framework of the Plan-TL.},
   author = {Asier Gutiérrez Fandiño and Jordi Armengol Estapé and Marc Pàmies and Joan Llop Palao and Joaquin Silveira Ocampo and Casimiro Pio Carrino and Carme Armentano Oller and Carlos Rodriguez Penagos and Aitor Gonzalez Agirre and Marta Villegas},
   doi = {10.26342/2022-68-3},
   issn = {1135-5948},
   journal = {Procesamiento del Lenguaje Natural},
   keywords = {Artificial intelligence,Benchmarking,Data processing.,MarIA,Natural language processing,Spanish language modelling,Spanish language resources,Tractament del llenguatge natural (Informàtica),Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Llenguatge natural},
   publisher = {Sociedad Española para el Procesamiento del Lenguaje Natural},
   title = {MarIA: Spanish Language Models},
   volume = {68},
   url = {https://upcommons.upc.edu/handle/2117/367156#.YyMTB4X9A-0.mendeley},
   year = {2022},
}

Disclaimer