gvic-unb/access-level-classification-brazilian-SEI

Access Level Classification of Brazilian SEI Documents

Dataset Summary

This dataset contains 606 documents in Brazilian Portuguese (and some English) labeled by access level, for training models that automatically classify documents processed through Brazil's Sistema Eletrônico de Informações (SEI) — the electronic document and process management system used across Brazilian federal public institutions.

When users upload external documents to SEI, they must manually assign one of three access levels: public, restricted, or confidential. This corpus was built to support an automatic classification approach that reduces human error in this labeling step.

The corpus combines real public SEI documents (collected via web crawling from Universidade de Brasília's — UnB — public SEI records) with artificially generated restricted and confidential samples (produced with ChatGPT from official UnB SEI form templates and public complaint text, with all personally identifiable information replaced by fictional data), since real restricted/confidential SEI documents cannot be publicly released.

This dataset accompanies the paper:

Borges, A. C. B., Marinho, M. C., Nogueira, R. F., Bordim, J. L., & Borges, V. R. P. (2025). Automatic Classification of Access Levels in Documents from the Brazilian Electronic Information System. Presented at the Congresso da Sociedade Brasileira de Computação (CSBC 2025), published in the Anais do Workshop de Computação Aplicada em Governo Eletrônico (WCGE). SBC. https://sol.sbc.org.br/index.php/wcge/article/view/36332

Code and templates used to build the corpus: https://gitlab.com/gvic-unb/classification-access-level-sei

Supported Tasks

• —Multi-class text classification: given a document's text, predict its access level (public, restricted, or confidential).

Languages

Documents are primarily in Brazilian Portuguese (pt-BR), with some samples in English.

Dataset Structure

Data Instances

Each row is one document. Example:

{
  "text": "Boletim de Atos Oficiais da UnB em 14/03/2025\n  ATO DA COORDENADORIA DE APOSENTADORIA E PENSÃO Nº 499/2025 ...",
  "label": 0
}

Data Fields

Class Labels and Distribution

• —Public documents were scraped directly from UnB's public SEI records (e.g., official bulletins, appointment/retirement acts).
• —Restricted documents were assembled from UnB SEI form templates (leave-of-absence requests, scholarship agreements, travel reports, etc.) filled with fictional data via ChatGPT, plus supporting documents (certificates, invoices, flight tickets) collected online with all identifying fields replaced.
• —Confidential documents simulate ombudsman (Ouvidoria) reports and complaints (e.g., harassment, workplace incidents) and medical/psychological reports, generated with ChatGPT or sourced from public complaint platforms, again with fictional data. This class is intentionally the smallest, reflecting the real-world proportion of confidential documents within SEI processes.

Data Splits

Released as a single split (train) with all 606 examples. No official train/test split is provided; the original paper used Stratified 5-Fold Cross-Validation. Note the corpus contains some duplicate/near-duplicate rows (a byproduct of generating multiple documents from the same form templates); consider deduplication depending on your use case.

Dataset Creation

Source Data

• —Public: extracted via a Selenium-based crawler from UnB's public SEI document search page.
• —Restricted: derived from 12 official UnB SEI form templates populated with fictional content by ChatGPT, plus internet-sourced supporting documents (certificates, utility bills, voter compliance certificates, etc.) OCR'd with Tesseract and anonymized.
• —Confidential: derived from ChatGPT-generated ombudsman complaints/reports (varying formal/informal writing styles) inspired by real complaint text from a Brazilian consumer-complaint platform, plus anonymized medical/psychological report templates filled with fictional patient data.

Annotations

Labels were assigned by construction (each document was collected or generated specifically to represent one access level), guided by domain experts on the SEI system at UnB and by Brazil's Freedom of Information Law (Lei de Acesso à Informação — LAI) and General Data Protection Law (LGPD).

Considerations for Using the Data

• —All restricted and confidential documents contain fictional, synthetically generated content — no real personal or confidential data is included, by design, for privacy/legal compliance (LAI and LGPD).
• —The dataset is small and imbalanced, especially for the confidential class (41 examples); the original paper reports this class as the hardest to classify.
• —Intended for research on access-level/privacy classification of administrative documents; not a substitute for official SEI access-level determinations.

Citation

If you use this dataset, please cite:

@inproceedings{borges2025access,
 author={Borges, Ana Clara B and Marinho, Mayara C and de Freitas Nogueira, Rodrigo and Bordim, Jacir L and Borges, Vinicius R P},
 title = { Automatic Classification of Access Levels in Documents from the Brazilian Electronic Information System},
 booktitle = {Anais do XIII Latin American Symposium on Digital Government},
 location = {Maceió/AL},
 year = {2025},
 keywords = {},
 issn = {2763-8723},
 pages = {203--214},
 publisher = {SBC},
 address = {Porto Alegre, RS, Brasil},
 doi = {10.5753/lasdigov.2025.9117},
 url = {https://sol.sbc.org.br/index.php/wcge/article/view/36332}
}

Authors

• —Ana Clara B. Borges — Faculdade de Ciências e Tecnologias em Engenharia, Universidade de Brasília
• —Mayara C. Marinho — Departamento de Ciência da Computação, Universidade de Brasília
• —Rodrigo de Freitas Nogueira — Arquivo Central, Universidade de Brasília
• —Jacir L. Bordim — Departamento de Ciência da Computação, Universidade de Brasília
• —Vinicius R. P. Borges — Departamento de Ciência da Computação, Universidade de Brasília

Universidade de Brasília (UnB), Brazil.

Source Repository

Original dataset and source code repository: CSBC 2025 – Automatic Classification of Access Levels in Documents from the Brazilian Electronic Information System, presented at the Congresso da Sociedade Brasileira de Computação (CSBC 2025).