robertuito-base-deacc

RoBERTuito

A pre-trained language model for social media text in Spanish

READ THE FULL PAPER Github Repository

RoBERTuito is a pre-trained language model for user-generated content in Spanish, trained following RoBERTa guidelines on 500 million tweets. RoBERTuito comes in 3 flavors: cased, uncased, and uncased+deaccented.

We tested RoBERTuito on a benchmark of tasks involving user-generated text in Spanish. It outperforms other pre-trained language models for this language such as BETO, BERTin and RoBERTa-BNE. The 4 tasks selected for evaluation were: Hate Speech Detection (using SemEval 2019 Task 5, HatEval dataset), Sentiment and Emotion Analysis (using TASS 2020 datasets), and Irony detection (using IrosVa 2019 dataset).

model	hate speech	sentiment analysis	emotion analysis	irony detection	score
robertuito-uncased	0.801 ± 0.010	0.707 ± 0.004	0.551 ± 0.011	0.736 ± 0.008	0.6987
robertuito-deacc	0.798 ± 0.008	0.702 ± 0.004	0.543 ± 0.015	0.740 ± 0.006	0.6958
robertuito-cased	0.790 ± 0.012	0.701 ± 0.012	0.519 ± 0.032	0.719 ± 0.023	0.6822
roberta-bne	0.766 ± 0.015	0.669 ± 0.006	0.533 ± 0.011	0.723 ± 0.017	0.6726
bertin	0.767 ± 0.005	0.665 ± 0.003	0.518 ± 0.012	0.716 ± 0.008	0.6666
beto-cased	0.768 ± 0.012	0.665 ± 0.004	0.521 ± 0.012	0.706 ± 0.007	0.6651
beto-uncased	0.757 ± 0.012	0.649 ± 0.005	0.521 ± 0.006	0.702 ± 0.008	0.6571

We release the pre-trained models on huggingface model hub:

Masked LM

To test the masked LM, take into account that space is encoded inside SentencePiece's tokens. So, if you want to test

Este es un día<mask>

don't put a space between día and <mask>

Usage

IMPORTANT -- READ THIS FIRST

RoBERTuito is not yet fully-integrated into huggingface/transformers. To use it, first install pysentimiento

pip install pysentimiento

and preprocess text using pysentimiento.preprocessing.preprocess_tweet before feeding it into the tokenizer

from transformers import AutoTokenizer
from pysentimiento.preprocessing import preprocess_tweet

tokenizer = AutoTokenizer.from_pretrained('pysentimiento/robertuito-base-cased')

text = "Esto es un tweet estoy usando #Robertuito @pysentimiento 🤣"
preprocessed_text = preprocess_tweet(text, ha)

tokenizer.tokenize(preprocessed_text)
# ['<s>','▁Esto','▁es','▁un','▁tweet','▁estoy','▁usando','▁','▁hashtag','▁','▁ro','bert','uito','▁@usuario','▁','▁emoji','▁cara','▁revolviéndose','▁de','▁la','▁risa','▁emoji','</s>']

We are working on integrating this preprocessing step into a Tokenizer within transformers library

Check a text classification example in this notebook:

Citation

If you use RoBERTuito, please cite our paper:

@inproceedings{perez-etal-2022-robertuito,
    title = "{R}o{BERT}uito: a pre-trained language model for social media text in {S}panish",
    author = "P{\'e}rez, Juan Manuel  and
      Furman, Dami{\'a}n Ariel  and
      Alonso Alemany, Laura  and
      Luque, Franco M.",
    booktitle = "Proceedings of the Thirteenth Language Resources and Evaluation Conference",
    month = jun,
    year = "2022",
    address = "Marseille, France",
    publisher = "European Language Resources Association",
    url = "https://aclanthology.org/2022.lrec-1.785",
    pages = "7235--7243",
    abstract = "Since BERT appeared, Transformer language models and transfer learning have become state-of-the-art for natural language processing tasks. Recently, some works geared towards pre-training specially-crafted models for particular domains, such as scientific papers, medical documents, user-generated texts, among others. These domain-specific models have been shown to improve performance significantly in most tasks; however, for languages other than English, such models are not widely available. In this work, we present RoBERTuito, a pre-trained language model for user-generated text in Spanish, trained on over 500 million tweets. Experiments on a benchmark of tasks involving user-generated text showed that RoBERTuito outperformed other pre-trained language models in Spanish. In addition to this, our model has some cross-lingual abilities, achieving top results for English-Spanish tasks of the Linguistic Code-Switching Evaluation benchmark (LinCE) and also competitive performance against monolingual models in English Twitter tasks. To facilitate further research, we make RoBERTuito publicly available at the HuggingFace model hub together with the dataset used to pre-train it.",
}