language:
- ca
license: apache-2.0
tags:
- catalan
- masked-lm
- longformer
- longformer-base-4096-ca-v2
- CaText
- Catalan Textual Corpus
widget:
- text: El Català és una llengua molt <mask>.
- text: Salvador Dalí va viure a <mask>.
- text: La Costa Brava té les millors <mask> d'Espanya.
- text: El cacaolat és un batut de <mask>.
- text: <mask> és la capital de la Garrotxa.
- text: Vaig al <mask> a buscar bolets.
- text: Antoni Gaudí vas ser un <mask> molt important per la ciutat.
- text: Catalunya és una referència en <mask> a nivell europeu.
Catalan Longformer (longformer-base-4096-ca-v2) base model
Table of Contents
Click to expand
Model description
The longformer-base-4096-ca-v2 is the Longformer version of the roberta-base-ca-v2 masked language model for the Catalan language. The use of these models allows us to process larger contexts (up to 4096 tokens) as input without the need of additional aggregation strategies. The pretraining process of this model started from the roberta-base-ca-v2 checkpoint and was pretrained for MLM on both short and long documents in Catalan.
The Longformer model uses a combination of sliding window (local) attention and global attention. Global attention is user-configured based on the task to allow the model to learn task-specific representations. Please refer to the original paper for more details on how to set global attention.
Intended uses and limitations
The longformer-base-4096-ca-v2 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.
How to use
Here is how to use this model:
from transformers import AutoModelForMaskedLM
from transformers import AutoTokenizer, FillMaskPipeline
from pprint import pprint
tokenizer_hf = AutoTokenizer.from_pretrained('projecte-aina/longformer-base-4096-ca-v2')
model = AutoModelForMaskedLM.from_pretrained('projecte-aina/longformer-base-4096-ca-v2')
model.eval()
pipeline = FillMaskPipeline(model, tokenizer_hf)
text = f"Em dic <mask>."
res_hf = pipeline(text)
pprint([r['token_str'] for r in res_hf])
Limitations and bias
At the time of submission, no measures have been taken to estimate the bias embedded in the model. However, we are well aware that our models may be biased since the corpus 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 training corpus consists of several corpora gathered from web crawling and public corpora.
| Corpus | Size in GB |
|---|---|
| Catalan Crawling | 13.00 |
| Wikipedia | 1.10 |
| DOGC | 0.78 |
| Catalan Open Subtitles | 0.02 |
| Catalan Oscar | 4.00 |
| CaWaC | 3.60 |
| Cat. General Crawling | 2.50 |
| Cat. Goverment Crawling | 0.24 |
| ACN | 0.42 |
| Padicat | 0.63 |
| RacoCatalá | 8.10 |
| Nació Digital | 0.42 |
| Vilaweb | 0.06 |
| Tweets | 0.02 |
For this specific pre-training process, we have performed an undersampling process to obtain a corpus of 5,3 GB.
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-base-bne pre-training consists of a masked language model training that follows the approach employed for the RoBERTa base. The training lasted a total of 37 hours with 8 computing nodes each one with 2 AMD MI50 GPUs of 32GB VRAM.
Evaluation
CLUB benchmark
The longformer-base-4096-ca-v2 model has been fine-tuned on the downstream tasks of the Catalan Language Understanding Evaluation benchmark (CLUB), that has been created along with the model.
It contains the following tasks and their related datasets:
Named Entity Recognition (NER)
NER (AnCora): extracted named entities from the original Ancora version, filtering out some unconventional ones, like book titles, and transcribed them into a standard CONLL-IOB format
Part-of-Speech Tagging (POS)
POS (AnCora): from the Universal Dependencies treebank of the well-known Ancora corpus.
Text Classification (TC)
TeCla: consisting of 137k news pieces from the Catalan News Agency (ACN) corpus, with 30 labels.
Textual Entailment (TE)
TE-ca: consisting of 21,163 pairs of premises and hypotheses, annotated according to the inference relation they have (implication, contradiction, or neutral), extracted from the Catalan Textual Corpus.
Semantic Textual Similarity (STS)
STS-ca: consisting of more than 3000 sentence pairs, annotated with the semantic similarity between them, scraped from the Catalan Textual Corpus.
Question Answering (QA):
VilaQuAD: contains 6,282 pairs of questions and answers, outsourced from 2095 Catalan language articles from VilaWeb newswire text.
ViquiQuAD: consisting of more than 15,000 questions outsourced from Catalan Wikipedia randomly chosen from a set of 596 articles that were originally written in Catalan.
CatalanQA: an aggregation of 2 previous datasets (VilaQuAD and ViquiQuAD), 21,427 pairs of Q/A balanced by type of question, containing one question and one answer per context, although the contexts can repeat multiple times.
Evaluation results
After fine-tuning the model on the downstream tasks, it achieved the following performance:
Evaluation results
| Task | NER (F1) | POS (F1) | STS-ca (Comb) | TeCla (Acc.) | TEca (Acc.) | VilaQuAD (F1/EM) | ViquiQuAD (F1/EM) | CatalanQA (F1/EM) | XQuAD-ca 1 (F1/EM) |
|---|---|---|---|---|---|---|---|---|---|
| RoBERTa-large-ca-v2 | 89.82 | 99.02 | 83.41 | 75.46 | 83.61 | 89.34/75.50 | 89.20/75.77 | 90.72/79.06 | 73.79/55.34 |
| RoBERTa-base-ca-v2 | 89.29 | 98.96 | 79.07 | 74.26 | 83.14 | 87.74/72.58 | 88.72/75.91 | 89.50/76.63 | 73.64/55.42 |
| Longformer-base-4096-ca-v2 | 88.49 | 98.98 | 78.37 | 73.79 | 83.89 | 87.59/72.33 | 88.70/76.05 | 89.33/77.03 | 73.09/54.83 |
| BERTa | 89.76 | 98.96 | 80.19 | 73.65 | 79.26 | 85.93/70.58 | 87.12/73.11 | 89.17/77.14 | 69.20/51.47 |
| mBERT | 86.87 | 98.83 | 74.26 | 69.90 | 74.63 | 82.78/67.33 | 86.89/73.53 | 86.90/74.19 | 68.79/50.80 |
| XLM-RoBERTa | 86.31 | 98.89 | 61.61 | 70.14 | 33.30 | 86.29/71.83 | 86.88/73.11 | 88.17/75.93 | 72.55/54.16 |
1 : Trained on CatalanQA, tested on XQuAD-ca.
Additional information
Author
Text Mining Unit (TeMU) at the Barcelona Supercomputing Center (bsc-temu@bsc.es)
Contact information
For further information, send an email to aina@bsc.es
Copyright
Copyright (c) 2022 Text Mining Unit at Barcelona Supercomputing Center
Licensing information
Funding
This work was funded by the Departament de la Vicepresidència i de Polítiques Digitals i Territori de la Generalitat de Catalunya within the framework of Projecte AINA.
Disclaimer
Click to expand
The models published in this repository are intended for a generalist purpose and are available to third parties. These models may have bias and/or any other undesirable distortions.
When third parties, deploy or provide systems and/or services to other parties using any of these models (or using systems based on these models) or become users of the models, they should note that it is their responsibility to mitigate the risks arising from their use and, in any event, to comply with applicable regulations, including regulations regarding the use of Artificial Intelligence.
In no event shall the owner and creator of the models (BSC – Barcelona Supercomputing Center) be liable for any results arising from the use made by third parties of these models.