BERTa: RoBERTa-based Catalan language model

Model description

BERTa is a transformer-based masked language model for the Catalan language.

It is based on the RoBERTA base model

and has been trained on a medium-size corpus collected from publicly available corpora and crawlers.

The pretrained model is available on the HuggingFace model hub under the name: https://huggingface.co/bsc/roberta-base-ca-cased

Training corpora and preprocessing

The training corpus consists of several corpora gathered from web crawling and public corpora.

The publicly available corpora are:

the Catalan part of the DOGC corpus, a set of documents from the Official Gazette of the Catalan Government
the Catalan Open Subtitles, a collection of translated movie subtitles
the non-shuffled version of the Catalan part of the OSCAR corpus \\cite{suarez2019asynchronous},

a collection of monolingual corpora, filtered from Common Crawl
The CaWac corpus, a web corpus of Catalan built from the .cat top-level-domain in late 2013

the non-deduplicated version
the Catalan Wikipedia articles downloaded on 18-08-2020.

The crawled corpora are:

The Catalan General Crawling, obtained by crawling the 500 most popular .cat and .ad domains
the Catalan Government Crawling, obtained by crawling the .gencat domain and subdomains, belonging to the Catalan Government
the ACN corpus with 220k news items from March 2015 until October 2020, crawled from the Catalan News Agency

To obtain a high-quality training corpus, each corpus have preprocessed with a pipeline of operations, including among the others,

sentence splitting, language detection, filtering of bad-formed sentences and deduplication of repetitive contents.

During the process, we keep document boundaries are kept.

Finally, the corpora are concatenated and further global deduplication among the corpora is applied.

The final training corpus consists of about 1,8B tokens.

Tokenization and pretraining

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 52,000 tokens.

The BERTa pretraining consists of a masked language model training that follows the approach employed for the RoBERTa base model

with the same hyperparameters as in the original work.

The training lasted a total of 48 hours with 16 NVIDIA V100 GPUs of 16GB DDRAM.

Downstream tasks

CLUB: Catalan Language Understanding Benchmark

The CLUB benchmark consists of 5 tasks, that are Part-of-Speech Tagging (POS), Named Entity Recognition (NER), Text Classification (TC), Semantic Textual Similarity (STS) and Question Answering (QA). For more information, refer to the HuggingFace datasets cards and Zenodo links below :

AnCora (POS):
- Splits info:
  - train: 13,123 examples
  - dev: 1,709 examples
  - test: 1,846 examples
- dataset card: https://huggingface.co/datasets/universal_dependencies
- data source: https://github.com/UniversalDependencies/UD_Catalan-AnCora
AnCora-ner (NER):
- Splits info:
  - train: 10,628 examples
  - dev: 1,427 examples
  - test: 1,526 examples
- dataset card: https://huggingface.co/datasets/bsc/ancora-ca-ner
- data source: https://zenodo.org/record/4762031#.YKaFjqGxWUk)
TeCla (TC):
- Splits info:
  - train: 110,203 examples
  - dev: 13,786 examples
  - test: 13,786 examples
- dataset card: https://huggingface.co/datasets/bsc/tecla
- data source: TeCla: consisting of 137k news pieces from the Catalan News Agency (ACN) corpus
STS-ca (STS):
- Splits info:
  - train: 2,073 examples
  - dev: 500 examples
  - test: 500 examples
- dataset card: https://huggingface.co/datasets/bsc/sts-ca
- data source: https://doi.org/10.5281/zenodo.4529183
ViquiQuAD (QA):
- Splits info:
  - train: 11,255 examples
  - dev: 1,492 examples
  - test: 1,429 examples
- dataset card: https://huggingface.co/datasets/bsc/viquiquad
- data source: https://doi.org/10.5281/zenodo.4562344
XQuAD (QA):
- Splits info:
  - test: 1,190 examples

dataset card: https://huggingface.co/datasets/bsc/xquad-ca
data source: https://doi.org/10.5281/zenodo.4526223

Fine-tuning and evaluation

The fine-tuning scripts for the downstream tasks are based on the HuggingFace Transformers library. To fine-tune and evaluate on CLUB, run the following commands:

The Catalan Language Understanding Benchmark (CLUB)

Fine-tune and evaluate on CLUB

To fine-tune and evaluate BERTa on the CLUB datasets, run the following commands:

bash setup_venv.sh
bash finetune_berta_club.sh

The commands above will run fine-tuning and evaluation on CLUB and the results will be shown in the results.json file. and the logs in the finetune_berta_club.log file.

Results

The official results obtained are:

Task	NER (F1)	POS (F1)	STS (Pearson)	TC (accuracy)	QA (ViquiQuAD) (F1/EM)	QA (XQuAD) (F1/EM)
BERTa	88.13	98.97	79.73	74.16	86.97/72.29	68.89/48.87
mBERT	86.38	98.82	76.34	70.56	86.97/72.22	67.15/46.51
XLM-RoBERTa	87.66	98.89	75.40	71.68	85.50/70.47	67.10/46.42
WikiBERT-ca	77.66	97.60	77.18	73.22	85.45/70.75	65.21/36.60

How to cite

If you use any of these resources in your work, please cite our latest paper:

@misc{armengolestape2021multilingual,
      title={Are Multilingual Models the Best Choice for Moderately Under-resourced Languages? A Comprehensive Assessment for Catalan}, 
      author={Jordi Armengol{-}Estap{\'{e}} and Casimiro Pio Carrino and Carlos Rodriguez-Penagos and Ona de Gibert Bonet and Carme Armentano{-}Oller and Aitor Gonzalez{-}Agirre and Maite Melero and Marta Villegas},
      year={2021},
      eprint={2107.07903},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

jordiluque/berta