PtBrVId

PtBrVId is a Portuguese Variety Identification corpus, built by combining pre-existing datasets originally created for different NLP tasks and released under permissive licenses.

Our goal is to provide a large, diverse, and multi-domain resource for studying and improving automatic identification of European Portuguese (PT-PT) and Brazilian Portuguese (PT-BR).

---

📚 Data Sources

The corpus is composed of datasets from various domains, each selected to ensure (as much as possible) mono-variety content. The current release is silver-labeled and unsupervised, meaning that we cannot fully guarantee that all documents are strictly mono-variety. A future version will include a refined annotation schema with both automatic and manual verification.

Domain	Variety	Dataset	Original Task	# Docs	License	Silver Labeled
Literature	PT-PT	Arquivo Pessoa	-	~4k	CC	✔
		Gutenberg Project	-	6	CC	✔
		LT-Corpus	-	56	ELRA END USER	✘
	PT-BR	Brazilian Literature	Author Identification	81	CC	✘
		LT-Corpus	-	8	ELRA END USER	✘
Politics	PT-PT	Koehn (2005) Europarl	Machine Translation	~10k	CC	✘
	PT-BR	Brazilian Senate Speeches1	-	~5k	CC	✔
Journalistic	PT-PT	CETEM Público	-	1M	CC	✘
	PT-BR	CETEM Folha	-	272k	CC	✘
Social Media	PT-PT	Ramalho (2021)	Fake News Detection	2M	MIT	✔
	PT-BR	Vargas (2022)	Hate Speech Detection	5k	CC-BY-NC-4.0	✘
		Cunha (2021)	Fake News Detection	2k	GPL-3.0	✔
Web	BOTH	Ortiz-Suarez (2020)	-	10k	CC	✔

Table 1: PtBrVId data sources and metadata.

¹ The Brazilian Senate Speeches dataset was created by the authors through web crawling of the Brazilian Senate website and is available on Hugging Face.

A raw version of the dataset is available here.

---

🛠 Annotation & Preprocessing

Annotation

We selected data sources known to contain primarily mono-variety Portuguese texts. While this approach helps ensure quality, this first release is entirely unsupervised. A planned v2 will introduce a hybrid annotation strategy combining automated labeling and manual review.

Preprocessing Pipeline

To standardize and clean the data, we applied the following steps:

1. Remove NaN values.
2. Remove empty documents.
3. Remove duplicate documents.
4. Apply the clean-text library to strip non-relevant content for variety identification.
5. Remove outlier documents with lengths below Q1 - 1.5 × IQR or above Q3 + 1.5 × IQR, where Q1 and Q3 are the first and third quartiles, and IQR is the interquartile range.

---

📖 Citation

If you use this corpus, please cite:

@article{Sousa_Almeida_Silvano_Cantante_Campos_Jorge_2025,
  title={Enhancing Portuguese Variety Identification with Cross-Domain Approaches},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={39},
  number={24},
  pages={25192--25200},
  year={2025},
  doi={10.1609/aaai.v39i24.34705},
  author={Sousa, Hugo and Almeida, Rúben and Silvano, Purificação and Cantante, Inês and Campos, Ricardo and Jorge, Alípio}
}