|
--- |
|
language: "ca" |
|
tags: |
|
- masked-lm |
|
- RoBERTa-base-ca-v2 |
|
- catalan |
|
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." |
|
license: apache-2.0 |
|
--- |
|
|
|
## Model description |
|
|
|
RoBERTa-ca-v2 is a transformer-based masked language model for the Catalan language. |
|
It is based on the [RoBERTA](https://github.com/pytorch/fairseq/tree/master/examples/roberta) base model |
|
and has been trained on a medium-size corpus collected from publicly available corpora and crawlers. |
|
|
|
## Tokenization and pretraining |
|
|
|
The training corpus has been tokenized using a byte version of [Byte-Pair Encoding (BPE)](https://github.com/openai/gpt-2) |
|
used in the original [RoBERTA](https://github.com/pytorch/fairseq/tree/master/examples/roberta) model with a vocabulary size of 52,000 tokens. |
|
The RoBERTa-ca-v2 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. |
|
|
|
## Training corpora and preprocessing |
|
|
|
The training corpus consists of several corpora gathered from web crawling and public corpora. |
|
|
|
|
|
| Corpus | Size in GB | |
|
|-------------------------|------------| |
|
| BNE-ca | 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 | |
|
|
|
## Evaluation |
|
|
|
### CLUB benchmark |
|
|
|
The BERTa 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: |
|
|
|
1. Part-of-Speech Tagging (POS) |
|
|
|
Catalan-Ancora: from the [Universal Dependencies treebank](https://github.com/UniversalDependencies/UD_Catalan-AnCora) of the well-known Ancora corpus |
|
|
|
2. Named Entity Recognition (NER) |
|
|
|
**[AnCora Catalan 2.0.0](https://zenodo.org/record/4762031#.YKaFjqGxWUk)**: extracted named entities from the original [Ancora](https://doi.org/10.5281/zenodo.4762030) version, |
|
filtering out some unconventional ones, like book titles, and transcribed them into a standard CONLL-IOB format |
|
|
|
3. Text Classification (TC) |
|
|
|
**[TeCla](https://doi.org/10.5281/zenodo.4627197)**: consisting of 137k news pieces from the Catalan News Agency ([ACN](https://www.acn.cat/)) corpus |
|
|
|
4. Semantic Textual Similarity (STS) |
|
|
|
**[Catalan semantic textual similarity](https://doi.org/10.5281/zenodo.4529183)**: consisting of more than 3000 sentence pairs, annotated with the semantic similarity between them, |
|
scraped from the [Catalan Textual Corpus](https://doi.org/10.5281/zenodo.4519349) |
|
|
|
5. Question Answering (QA): |
|
|
|
**[ViquiQuAD](https://doi.org/10.5281/zenodo.4562344)**: 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. |
|
|
|
**[XQuAD](https://doi.org/10.5281/zenodo.4526223)**: the Catalan translation of XQuAD, a multilingual collection of manual translations of 1,190 question-answer pairs from English Wikipedia used only as a _test set_ |
|
|
|
Here are the train/dev/test splits of the datasets: |
|
|
|
| Task (Dataset) | Total | Train | Dev | Test | |
|
|:--|:--|:--|:--|:--| |
|
| NER (Ancora) |13,581 | 10,628 | 1,427 | 1,526 | |
|
| POS (Ancora)| 16,678 | 13,123 | 1,709 | 1,846 | |
|
| STS | 3,073 | 2,073 | 500 | 500 | |
|
| TC (TeCla) | 137,775 | 110,203 | 13,786 | 13,786| |
|
| QA (ViquiQuAD) | 14,239 | 11,255 | 1,492 | 1,429 | |
|
|
|
### Results |
|
|
|
| Task | NER (F1) | POS (F1) | STS (Pearson) | TC (accuracy) | QA (ViquiQuAD) (F1/EM) | QA (XQuAD) (F1/EM) | |
|
| ------------|:-------------:| -----:|:------|:-------|:------|:----| |
|
| RoBERTa-base-ca-v2 | **89.84** | **99.07** | **79.98** | **83.41** | **88.04/74.65** | **71.50/53.41** | |
|
| 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 | |
|
|
|
## Intended uses & limitations |
|
The model is ready-to-use only for masked language modelling to perform the Fill Mask task (try the inference API or read the next section) |
|
However, the is intended to be fine-tuned on non-generative downstream tasks such as Question Answering, Text Classification or Named Entity Recognition. |
|
|
|
|
|
## Funding |
|
This work was funded by the Generalitat de Catalunya within the framework of the AINA language technologies plan. |
|
|