CLUB: Catalan Language Understanding Benchmark

Tasks and datasets

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 :

1. AnCora (POS):

   • Splits info:

     • train: 13,123 examples
     • validation: 1,709 examples
     • test: 1,846 examples

   • dataset card: https://huggingface.co/datasets/universal_dependencies

   • data source: https://github.com/UniversalDependencies/UD_Catalan-AnCora

2. AnCora-ner (NER):

   • Splits info:

     • train: 10,628 examples
     • validation: 1,427 examples
     • test: 1,526 examples

   • dataset card: https://huggingface.co/datasets/projecte-aina/ancora-ca-ner

   • data source: https://zenodo.org/record/4762031#.YKaFjqGxWUk)

3. TeCla (TC):

   • Splits info:

     • train: 110,203 examples
     • validation: 13,786 examples
     • test: 13,786 examples

   • dataset card: https://huggingface.co/datasets/projecte-aina/tecla

   • data source: TeCla: consisting of 137k news pieces from the Catalan News Agency (ACN) corpus

4. STS-ca (STS):

   • Splits info:

     • train: 2,073 examples
     • validation: 500 examples
     • test: 500 examples

   • dataset card: https://huggingface.co/datasets/projecte-aina/sts-ca

   • data source: https://doi.org/10.5281/zenodo.4529183

5. ViquiQuAD (QA):

   • Splits info:

     • train: 11,255 examples
     • validation: 1,492 examples
     • test: 1,429 examples

   • dataset card: https://huggingface.co/datasets/projecte-aina

   • data source: https://doi.org/10.5281/zenodo.4562344

6. XQuAD (QA):

   • Splits info:
     • test: 1,190 examples

• dataset card: https://huggingface.co/datasets/projecte-aina/xquad-ca

• data source: https://doi.org/10.5281/zenodo.4526223

7. TECA (Textual Entailment)

   • Splits info:

     • train: 16,930 examples
     • validation: 2116
     • test: 2117

   • dataset card: https://huggingface.co/datasets/projecte-aina/teca

   • data source: https://zenodo.org/record10.5281/zenodo.4593271.

8. CaSum (text summarization)

   • Splits info:

     • train: 197,735 examples
     • validation: 10,000 examples
     • test: 10,000 examples

   • dataset card: https://huggingface.co/datasets/projecte-aina/casum

9. VilaSum (text summarization)

   • Splits info:

     • train: 13,843 examples

   • dataset card: https://huggingface.co/datasets/projecte-aina/vilasum

BERTa

BERTa is a transformer-based masked language model for the Catalan language, based on the RoBERTA base model

Pretrained model: https://huggingface.co/PlanTL-GOB-ES/roberta-base-ca <br/> Training corpora: https://doi.org/10.5281/zenodo.4519348

Fine-tune and evaluate on CLUB

To fine-tune and evaluate your model on the CLUB benchamark, run the following commands:

bash setup_venv.sh
bash run_club.sh <model_name_on_HF>

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

Fine-tuning and evaluation

For each model we used the same fine-tuning setting across tasks, consisting of 10 training epochs, with an effective batch size of 32 instances, a max input length of 512 tokens (128 tokens in the case of Textual Entailment though) and a learning rate of 5e−5. The rest of the hyperparameters are set to the default values in Huggingface Transformers scripts. We then select the best checkpoint as the one that maximised the task-specific metric on the corresponding validation set, and finally evaluate it on the test set.

Results

Evaluations results obtained running the scripts above with the <model_name_on_HF> set to PlanTL-GOB-ES/roberta-base-ca:

| Model | NER (F1) | POS (F1) | STS (Pearson) | TC (accuracy) | QA (ViquiQuAD) (F1/EM) | QA (XQuAD) (F1/EM) | TE (TECA) (accuracy) | | ------------|:-------------:| -----:|:------|:-------|:------|:----|:----| | BERTa | 89.63 | 98.93 | 81.20 | 74.04 | 86.99/73.25 | 67.81/49.43 | 79.12 | | mBERT | 86.38 | 98.82 | 76.34 | 70.56 | 86.97/72.22 | 67.15/46.51 | 74.78 | | XLM-RoBERTa | 87.66 | 98.89 | 75.40 | 71.68 | 85.50/70.47 | 67.10/46.42 | 75.44 | | WikiBERT-ca | 77.66 | 97.60 | 77.18 | 73.22 | 85.45/70.75 | 65.21/36.60 | x |

How to cite

If you use any of these resources (datasets or models) in your work, please cite our latest paper:

@inproceedings{armengol-estape-etal-2021-multilingual,
    title = "Are Multilingual Models the Best Choice for Moderately Under-resourced Languages? {A} Comprehensive Assessment for {C}atalan",
    author = "Armengol-Estap{\'e}, Jordi  and
      Carrino, Casimiro Pio  and
      Rodriguez-Penagos, Carlos  and
      de Gibert Bonet, Ona  and
      Armentano-Oller, Carme  and
      Gonzalez-Agirre, Aitor  and
      Melero, Maite  and
      Villegas, Marta",
    booktitle = "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.findings-acl.437",
    doi = "10.18653/v1/2021.findings-acl.437",
    pages = "4933--4946",
}