gvic-unb/extracurricular-certificates-declarations

Extracurricular Certificates and Declarations Classification

Dataset Summary

This dataset contains 603 academic proof documents (certificates, declarations, scientific papers, and other academic documents) in Brazilian Portuguese and English, labeled according to the pre-defined groups of the extracurricular activities accreditation regulation of the Computer Science undergraduate program at Universidade de Brasília (UnB).

Brazilian undergraduate programs allow students to accredit extracurricular activities (conferences, research programs, tutoring, publications, etc.) in their academic transcripts by submitting proof documents. This corpus was built to train NLP/ML models that automatically classify these documents into the correct accreditation group, reducing the manual, time-consuming, and error-prone work of academic staff.

This dataset accompanies the paper:

Cavalcante, J. P. F. M., Marinho, M. C., & Borges, V. R. P. (2024). Natural Language Processing Approaches for Accrediting Students on Extracurricular Activities. Anais do XXXV Simpósio Brasileiro de Informática na Educação (SBIE 2024), pages 1796-1809. SBC. https://doi.org/10.5753/sbie.2024.242548

Original repository (code + corpus): you can find here.

Supported Tasks

• Multi-class text classification: given the raw text extracted from a proof document, predict which extracurricular-activity group it belongs to (or whether it's invalid).

Languages

Documents are in Brazilian Portuguese (pt-BR) and English, reflecting real certificates/declarations collected from national and international sources.

Dataset Structure

Data Instances

Each row is one document. Example:

{
  "text": "Analise da Mobilidade Urbana por meio de Redes Sociais\nBaseadas em Localizacao Estudo de Caso em Cidades Inteligentes\nJoaquim B S Ferreira ...",
  "label": 1
}

Data Fields

Field	Type	Description
text	string	Raw text extracted from the proof document (via PyMuPDF for PDFs, Tesseract OCR for images)
label	int	Accreditation category (see Class Labels below)

Class Labels

The label values are the raw category codes used during annotation (LabelStudio); they do not run sequentially from 1-8 as in the paper's "Group" numbering, but correspond 1-to-1 by matching instance counts against Table 3 of the paper:

label	Group (paper)	Description	Required proof document	# Instances	%
-1	—	Invalid / does not satisfy accreditation rules	—	66	11%
1	Group 1	Publication of scientific papers (conferences, workshops, journals)	Copy of the paper	92	15%
2	Group 2	Undergraduate Research Program (Iniciação Científica)	Certificate from the program or declaration signed by supervisor	73	12%
6	Group 3	Tutorial Education Program (PET)	Declaration issued by chancellors or professor	14	2%
8	Group 4	Lecturer / organizer / speaker / exhibitor at CS-related events or short courses	Certificate issued by the organization	41	7%
9	Group 5	Member of an undergraduate tutoring project	Declaration issued by chancellors or responsible professor	35	6%
10	Group 6	Participation in conferences, workshops, symposiums, short courses, meetings	Certificate issued by the organization	198	33%
11	Group 7	Team member of an extension project	Declaration issued by chancellors or responsible professor	34	6%
12	Group 8	Team member of a junior enterprise in computing	Declaration issued by junior enterprise leader or responsible professor	50	8%

Note: the corpus intentionally over-represents minority classes relative to their real-world frequency (see the "Number of instances per category in specialist's sample" table in the paper) to mitigate class imbalance during model training.

Data Splits

Released as a single split (train) with all 603 examples. No official train/test split is provided; the original paper used Stratified 5-Fold Cross-Validation. The corpus contains 1 duplicate row.

Dataset Creation

Source Data

• Certificates and declarations related to Computer Science were collected from image search engines (focusing on online courses) and from conference/symposium/event websites, national and international.
• Some certificates/declarations were artificially generated from official UnB templates (for internal events, extension projects, tutorial programs) with fictional names, events, and dates, since real institutional certificates could not be publicly collected.
• Text was extracted using PyMuPDF (PDF files) and Tesseract OCR (image files).

Annotations

Labeled by a specialist using the LabelStudio platform, following the accreditation rules (Table 1 in the paper) for the Computer Science undergraduate program at UnB. Documents not satisfying any predefined category were labeled -1 (e.g., a "Best Paper" certificate that doesn't show the student's name as speaker/presenter or the workload hours).

Considerations for Using the Data

• The dataset is small and originally imbalanced; classes -1, 3 (label 6), and 4 (label 8) had the fewest samples and were the hardest to classify in the original study (Macro F1 as low as 0.77 for class -1 with the best model).
• Some documents were synthetically generated with fictional data to avoid exposing real institutional/personal information.
• Labels reflect UnB's Computer Science program regulation specifically; categories may not generalize directly to other institutions' accreditation rules.

Citation

If you use this dataset, please cite:

@inproceedings{cavalcante2024nlp,
  author={Cavalcante, João Pedro F. M. and Marinho, Mayara C. and Borges, Vinicius R. P.},
  title={Natural Language Processing Approaches for Accrediting Students on Extracurricular Activities},
  booktitle={Anais do XXXV Simpósio Brasileiro de Informática na Educação (SBIE 2024)},
  year={2024},
  pages={1796--1809},
  publisher={SBC},
  doi={10.5753/sbie.2024.242548},
  url={https://sol.sbc.org.br/index.php/sbie/article/view/31358}
}

Funding

This research received funding from Fundação de Apoio à Pesquisa do Distrito Federal (FAPDF), Finance Code 10422B.

Authors

• João Pedro F. M. Cavalcante — Departamento de Estatística, Universidade de Brasília
• Mayara C. Marinho — 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.