Edit model card

Catalan BERTa (RoBERTa-base) finetuned for Semantic Textual Similarity.

The roberta-base-ca-cased-sts is a Semantic Textual Similarity (STS) model for the Catalan language fine-tuned from the BERTa model, a RoBERTa base model pre-trained on a medium-size corpus collected from publicly available corpora and crawlers (check the BERTa model card for more details).


We used the STS dataset in Catalan called STS-ca for training and evaluation.

Evaluation and results

We evaluated the roberta-base-ca-cased-sts on the STS-ca test set against standard multilingual and monolingual baselines:

Model STS-ca (Pearson)
roberta-base-ca-cased-sts 79.73
mBERT 76.34
XLM-RoBERTa 75.40
WikiBERT-ca 77.18

For more details, check the fine-tuning and evaluation scripts in the official GitHub repository.

How to use

To get the correct1 model's prediction scores with values between 0.0 and 5.0, use the following code:

from transformers import pipeline, AutoTokenizer
from scipy.special import logit

model = 'projecte-aina/roberta-base-ca-cased-sts'
tokenizer = AutoTokenizer.from_pretrained(model)
pipe = pipeline('text-classification', model=model, tokenizer=tokenizer)

def prepare(sentence_pairs):
    sentence_pairs_prep = []
    for s1, s2 in sentence_pairs:
        sentence_pairs_prep.append(f"{tokenizer.cls_token} {s1}{tokenizer.sep_token}{tokenizer.sep_token} {s2}{tokenizer.sep_token}")
    return sentence_pairs_prep

sentence_pairs = [("El llibre va caure per la finestra.", "El llibre va sortir volant."),
                  ("M'agrades.", "T'estimo."),
                  ("M'agrada el sol i la calor", "A la Garrotxa plou molt.")]

predictions = pipe(prepare(sentence_pairs), add_special_tokens=False)

# convert back to scores to the original 1 and 5 interval
for prediction in predictions:
    prediction['score'] = logit(prediction['score'])

Expected output:

[{'label': 'SIMILARITY', 'score': 2.4280577200108384}, 
{'label': 'SIMILARITY', 'score': 2.132843521240822}, 
{'label': 'SIMILARITY', 'score': 1.615101695426227}]

1 avoid using the widget scores since they are normalized and do not reflect the original annotation values.


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

    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",
Downloads last month
Hosted inference API
This model can be loaded on the Inference API on-demand.

Dataset used to train projecte-aina/roberta-base-ca-cased-sts

Evaluation results