RelEx-PT

Dataset Summary

RelEx-PT, a new sentence-level Relation Extraction dataset for Portuguese. Addressing the scarcity of high-quality, controlled resources for the language, RelEx-PT provides a balanced benchmark comprising 18 Wikidata-derived relation types across diverse domains. The dataset is built through a distant supervision pipeline that links Wikidata triples with Portuguese Wikipedia sentences and enhanced by a Natural Language Inference (NLI)-based filtering process, combining scalability with quality assurance.

Dataset Creation

For full construction pipeline and source code check: https://github.com/TomasCCPinto/RelEx-PT

Licensing Information

RelEx-PT is released under the Creative Commons Attribution-ShareAlike 4.0 International (CC BY-SA 4.0) license.

Citation

If you use RelEx-PT in your research, please cite:

@inproceedings{pinto-etal-2026-relex,
  title = {RelEx-PT: A Portuguese Sentence-Level Relation Extraction Dataset},
  author = {Pinto, Tomás and Silva, Catarina and Oliveira, Hugo Goncalo},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference (LREC 2026)},
  month = {May},
  year = {2026},
  pages = {7681--7691},
  address = {Palma, Mallorca, Spain},
  publisher = {European Language Resources Association (ELRA)},
  editor = {Piperidis, Stelios and Bel, Núria and van den Heuvel, Henk and Ide, Nancy and Krek, Simon and Toral, Antonio},
  doi = {10.63317/473a5ncxrap4},
  abstract = {We introduce RelEx-PT, a new sentence-level Relation Extraction dataset for Portuguese. Addressing the scarcity of high-quality, controlled resources for the language, RelEx-PT provides a balanced benchmark comprising 18 Wikidata-derived relation types across diverse domains. The dataset is built through a distant supervision pipeline that links Wikidata triples with Portuguese Wikipedia sentences and enhanced by a Natural Language Inference (NLI)-based filtering process, combining scalability with quality assurance. Additionally, we conduct baseline experiments to evaluate the dataset’s applicability across diverse extraction settings, including Relation Classification (RC), Relation Triple Extraction, and Open Information Extraction. These experiments leverage both prompting and fine-tuning strategies using Large Language Models. The results show that RelEx-PT effectively supports a range of extraction paradigms, yielding high performance in RC and competitive results in structured triple generation, while also highlighting key challenges in open-ended extraction.}
}