Flaglab/econ-ie-spanish-ner

Econ-IE (Spanish NER)

The Econ-IE (Spanish) dataset is a translated and cleaned version of the original EconBERTA corpus, which introduced a Named Entity Recognition (NER) benchmark focused on impact evaluation in economics.
The Spanish version was created to enable benchmarking of domain-specific NER models for Spanish scientific and policy-oriented text.

---

Dataset Summary

The original EconBERTA dataset was developed entirely in English.
To make this resource available to the Spanish NLP community, we translated the entire corpus into Spanish using the GPT-4o model.

The translation process was carefully designed to preserve the BIO tagging scheme and all original entity labels.
A custom prompt was engineered to ensure that tokens were translated while maintaining the same structure and entity consistency:

prompt = f"""
You are an expert economic translator. Translate each numbered sequence into Spanish,
keeping the EXACT same labels and the BIO tagging scheme. If a token expands into multiple
words, the first subtoken inherits B-XXX and the following ones inherit I-XXX.
Do not remove or modify any LABEL.

{payload}

Respond in the same numbered format, as follows:
1. translation-of-sentence-1
2. translation-of-sentence-2
...
without any additional explanations.
""".strip()

---

Data Creation and Cleaning

After several iterations and prompt refinements, the translation process was automated for the entire dataset.
During early experiments, common translation errors were identified, including:

• Missing labels when tokens expanded into multiple words,
• Introduction of spurious entities, and
• Misassigned BIO tags on tokens labeled as "O".

To correct these issues, post-processing scripts were developed to enforce consistent label propagation and ensure alignment between tokens and entity tags.
Additionally, a manual review of a random subset of translated samples was conducted to verify translation accuracy and label consistency.

During the manual validation phase, we also detected inconsistencies in the original English dataset hosted in the EconBERTA repository:

• The development split contained samples duplicated from the training set.
• All three splits included repeated sentences, sometimes up to four times.

We therefore applied a full deduplication and integrity check on the English corpus and the translated Spanish version.
After cleaning, new stratified splits were created to preserve entity distribution across partitions.

---

Dataset Statistics

Split	# Sentences
Train	5,268
Validation	1,129
Test	1,129
Total	7,526

Each sample consists of a tokenized sentence and a corresponding list of BIO-formatted NER labels.

---

Data Fields

Field	Type	Description
tokens	list[string]	Tokenized words of the sentence.
ner_tags	list[int]	List of integer-encoded entity tags following the BIO scheme.

Entity tag mapping is stored in the dataset’s metadata (ClassLabel) and can be retrieved programmatically:

from datasets import load_dataset
ds = load_dataset("FlagLab/econ-ie-spanish-ner")
labels = ds["train"].features["ner_tags"].feature.names
print(labels)

---

Languages

• Spanish (es)

---

Use Cases

• Training and evaluation of Named Entity Recognition (NER) models in the economic and social policy domains.
• Cross-lingual transfer experiments from English to Spanish using domain-specific corpora.
• Fine-tuning and benchmarking of scientific Spanish encoders such as Sci-BETO or Sci-RoBERTa.

---

License

Creative Commons Attribution 4.0 International (CC BY 4.0)
Users must attribute the original EconBERTA dataset and this Spanish adaptation when redistributing or using the data for research.

Full license text: https://creativecommons.org/licenses/by/4.0/

---

Acknowledgments

We thank the authors of the EconBERTA paper for making the original dataset openly available, and the broader NLP community for supporting open-source resources.