license: mit
language:
- pt
- en
pretty_name: UFLA-FORMS
task_categories:
- token-classification
- question-answering
- table-question-answering
configs:
- config_name: 10-fold
data_files:
- split: '0'
path: 10-fold/0-*
- split: '1'
path: 10-fold/1-*
- split: '2'
path: 10-fold/2-*
- split: '3'
path: 10-fold/3-*
- split: '4'
path: 10-fold/4-*
- split: '5'
path: 10-fold/5-*
- split: '6'
path: 10-fold/6-*
- split: '7'
path: 10-fold/7-*
- split: '8'
path: 10-fold/8-*
- split: '9'
path: 10-fold/9-*
- config_name: train_test
data_files:
- split: train
path: train_test/train-*
- split: test
path: train_test/test-*
default: true
- config_name: train_validation_test
data_files:
- split: train
path: train_validation_test/train-*
- split: validation
path: train_validation_test/validation-*
- split: test
path: train_validation_test/test-*
UFLA-FORMS: an Academic Forms Dataset for Information Extraction in the Portuguese Language
About
UFLA-FORMS is a manually labeled dataset of document forms in Brazilian Portuguese extracted from the domains of the Federal University of Lavras (UFLA). The dataset emphasizes the hierarchical structure between the entities of a document through their relationships, in addition to the extraction of key-value pairs. Samples were labeled using ToolRI.
Overview
UFLA-FORMS comprises of 200 labeled samples containing 7,710 entities with 34,396 words and 4,442 relationships between them. Each sample is composed of a PNG image extracted from a filled form and a JSON file with it's metadata. The metadata covers labeling information such as sentences, bounding boxes, class labels and relationship pairs. Each entity of the samples were labeled into one of the four classes, according to the FUNSD dataset: Header, Question, Answer or Other. Images are available in the image folder and metadata in the data folder. Each sample has the same name (ranging from "000" to "199") for its image and respective metadata in the files names before extensions. A dataset_info.json file is available containing information about the dataset, including it's partitions defined for cross-validation and hold-out experiments.
Examples
The images below presents examples of samples from UFLA-FORMS. The highlight colors yellow, blue, green and red represent, respectively, entities of the classes Header, Question, Answer and Other.
Download
UFLA-FORMS is available in its original format in its repository.
Citation
If you find UFLA-FORMS useful in your research, please cite our work:
@article{uflaforms,
author={Victor Gonçalves Lima and Denilson Alves Pereira},
title={{UFLA-FORMS: an Academic Forms Dataset for Information Extraction in the Portuguese Language}},
journal={Language Resources and Evaluation},
publisher={Springer},
doi={10.1007/s10579-024-09802-3},
url={https://doi.org/10.1007/s10579-024-09802-3},
misc={PDF: https://rdcu.be/d69VH, dataset: https://github.com/LabRI-Information-Retrieval-Lab/UFLA-FORMS/},
year={2025}
}