JusBR-ER

The public benchmark repository for JusBR-ER, a benchmark for entity resolution of legal concepts in Brazilian Portuguese.

JusBR-ER contains 1,071 labeled pairs of Brazilian legal concepts validated by a practicing lawyer (OAB/PR 96,036), covering synonyms, writing variants, antonym traps, and genus-species distinctions.

Language: pt-BR | English

Repository Snapshot

Field	Value
Task	Entity resolution of legal concepts
Language	Brazilian Portuguese
Total pairs	1,071
Positive pairs	530 merge
Negative pairs	541 no_merge
Validator	Practicing lawyer, OAB/PR 96,036
Version	1.0

Repository Contents

Asset	Description
jusbr_er_v1.json	Main benchmark dataset
synonym_dict_full.json	Validated synonym dictionary
synonym_dict_sample.json	Sample subset of synonym pairs
eval_baseline.py	Standalone baseline evaluation script
annotation_guidelines.md	Annotation guidelines and borderline rulings

What The Benchmark Covers

JusBR-ER is designed for legal concept matching, not just generic string similarity.

It includes:

• technical synonyms
• writing variants
• legal concepts that look similar but should stay separate
• antonym and opposition traps
• genus-species and sibling-category distinctions

Dataset Format

{
  "metadata": {
    "name": "JusBR-ER",
    "version": "1.0"
  },
  "pairs": [
    {
      "pair_id": 87,
      "entity_a": "Periculum in mora",
      "entity_b": "Perigo de dano",
      "decision": "merge",
      "relationship_type": "sinonimo"
    }
  ]
}

Label Taxonomy

Type (pt-BR)	English gloss	Merge?
identico	identical	Yes
sinonimo	synonym	Yes
variante	writing variant	Yes
genero_especie	genus-species	No
especies_mesmo_genero	same-genus siblings	No
antonimo_complementar	complementary antonym	No
sem_relacao	unrelated	No

Baseline Snapshot

Setting	Precision	Recall	F1
JW threshold 0.95	92.1%	10.9%	19.6%
JW threshold 0.88 (default)	50.2%	24.1%	32.6%
JW threshold 0.70	55.6%	81.9%	66.3%
Embedding-first blocking + Claude Sonnet 4.6	-	-	85.9%

Quick Start

Run the standalone baseline:

python eval_baseline.py --dataset jusbr_er_v1.json

Run the threshold sweep:

python eval_baseline.py --dataset jusbr_er_v1.json --sweep

No external dependencies are required. Python 3.8+ is enough.

Usage And Licensing

The repository uses a split licensing model:

• dataset and annotation guidelines: CC BY-NC-SA 4.0
• synonym dictionary: CC BY-NC-SA 4.0
• evaluation code: MIT

Free for academic and research use. Commercial use requires authorization: diego@sens.legal

Citation

@inproceedings{sens2026jusbrer,
  title={JusBR-ER: A Benchmark for Entity Resolution of Legal Concepts in Brazilian Portuguese},
  author={Sens, Diego},
  booktitle={Proceedings of the Natural Legal Language Processing Workshop (NLLP)},
  year={2026},
  url={https://github.com/sensdiego/jusbr-er}
}

Contact

Diego Sens — diego@sens.legal — sens.legal