piimb/mapa-eur-lex-pii

Dataset Card for MAPA EUR-LEX PII

A small, multilingual PII-detection benchmark built from the dglover1/mapa-eur-lex test split (itself derived from joelniklaus/mapa). Two English EUR-LEX legal documents were re-annotated by hand with a fine-grained PII scheme, and those labels were then projected onto their professional human translations in 20 other EU languages.

Dataset Details

Dataset Description

The dataset contains 42 documents: two source legal opinions, each available in 21 parallel language versions. Every document carries character-level PII entity spans. The English versions are hand-annotated; the other 20 languages are labelled by projecting the English annotations onto the existing human translations with an LLM.

It is intended as an evaluation set for multilingual PII detection and masking. The fixed two-document, 21-language structure makes it a parallel corpus: the same PII values can be compared across all languages.

• Curated by: camillevanhoffelen
• Language(s) (NLP): Bulgarian (bg), Czech (cs), Danish (da), German (de), Greek (el), English (en), Spanish (es), Estonian (et), Finnish (fi), French (fr), Irish (ga), Hungarian (hu), Italian (it), Lithuanian (lt), Latvian (lv), Maltese (mt), Dutch (nl), Portuguese (pt), Romanian (ro), Slovak (sk), Swedish (sv)
• License: CC BY 4.0

Dataset Sources

• Repository: dglover1/mapa-eur-lex, joelniklaus/mapa
• Paper: Spanish Datasets for Sensitive Entity Detection in the Legal Domain (de Gibert Bonet et al., LREC 2022)

Uses

Direct Use

Evaluating PII detection and masking systems, particularly multilingual ones, on legal-domain text. Because every PII value is present in all 21 languages, the dataset also supports cross-lingual comparison of detection quality.

Out-of-Scope Use

The dataset is small (42 documents from 2 source texts) and is meant for evaluation only — not for training. It is not a representative sample of legal text or of PII distributions in general.

Dataset Structure

The dataset is a single test.jsonl file with one document per line. Each row is a DocumentExample:

Field	Type	Description
uid	string	Document identifier (e.g. 205542_en, 205542_fr)
text	string	Full document text
language	string	ISO 639-1 language code
entities	list of {start, end, label}	Character-level PII spans (half-open [start, end))

The schema matches the DocumentExample model used by the piimb PII-masking benchmark.

Label set

The manual relabeling replaced the original MAPA coarse labels (PERSON, DATE, ADDRESS, ORGANISATION, AMOUNT) with a finer-grained PII scheme. Entity counts across all 42 documents:

Label	Count
unique_id	3929
date	1804
last_name	990
occupation	779
education_level	270
title	187
employer	187
first_name	171
age	63
url	21
race_ethnicity	21
nationality	20
Total	8442

Dataset Creation

Curation Rationale

The goal was a multilingual, document-level PII benchmark with a label scheme aligned to piimb, reusing MAPA's parallel translations rather than re-translating. The English documents were re-annotated from scratch because the source MAPA labels are unsuitable for PII-masking evaluation:

• Inconsistent — equivalent entities are tagged unevenly (e.g. some last names are omitted).
• Incompatible granularities — fine_grained is not a subset of coarse_grained, so neither can be derived from the other.
• Misleading type names — coarse names confuse zero-shot models like GLiNER (e.g. countries such as "Ireland" tagged ADDRESS).
• Poor coverage, over-broad types — relevant PII is left untagged, and some types are too broad to mask usefully (e.g. ORGANISATION covering "the WRC").

Source Data

Data Collection and Processing

The processing is implemented as a four-step CLI in this repository. The commands are registered under mapa-eur-lex-pii:

1. prepare-dataset — Load the dglover1/mapa-eur-lex test split, convert each sentence's coarse_grained BIO tags into character-level entity spans, drop the original tag columns, and write sentence-level rows to data/mapa-eur-lex.jsonl.

   uv run mapa-eur-lex-pii prepare-dataset
   

2. label-studio — Group the sentence rows back into documents (by file_name), shifting entity offsets accordingly, and emit a Label Studio import file (data/mapa-eur-lex-label-studio-export.json) plus a labeling interface config (data/labeling_config.xml). The source spans are pre-loaded as predictions so they can be reviewed and corrected. Pass --language en to export only the English documents.

   uv run mapa-eur-lex-pii label-studio
   

3. Manual relabeling — In Label Studio, the two English documents are re-annotated by hand with the fine-grained PII scheme above. The corrected English annotations are exported to data/mapa-eur-lex-annotations-en.json.

4. translate-labels — Project the hand-annotated English spans onto the 20 parallel translations of each document. For every target translation, a single LLM call (default claude-haiku-4-5, via litellm) re-emits the translation verbatim with inline <pii label="...">…</pii> tags inserted; the tags are parsed back into spans and realigned onto the authoritative target text. The result is written to test.jsonl as DocumentExample rows.

   uv run mapa-eur-lex-pii translate-labels
   

Who are the source data producers?

The underlying texts and their translations are EUR-LEX legal documents produced and professionally translated by the institutions of the European Union, redistributed via the MAPA project and the joelniklaus/mapa / dglover1/mapa-eur-lex datasets.

Annotations

Annotation process

English documents are annotated by a human in Label Studio. The non-English documents are not annotated directly: their labels are projected from the English annotations onto the existing human translations by an LLM, then offset-realigned with difflib.

Who are the annotators?

The English spans are annotated by camillevanhoffelen. The non-English spans are generated by an LLM (claude-haiku-4-5 by default) and have not been exhaustively human-verified.

Personal and Sensitive Information

The dataset is about personal data: it labels names, dates, occupations, identifiers, and similar PII for detection research. All content originates from published EUR-LEX legal documents.

Bias, Risks, and Limitations

• Small and narrow. Only 2 source documents in the legal domain; label counts are dominated by unique_id and date, and several labels (url, race_ethnicity, nationality) are rare.
• Machine-projected labels. For the 20 non-English languages, spans come from LLM projection and may contain missing, spurious, or misaligned entities, especially where a translation repeats a value in multiple forms.
• Not a training set. The size and provenance make it suitable for evaluation only.

Citation

@article{DeGibertBonet2022,
  author  = {{de Gibert Bonet}, Ona and {Garc{\'{i}}a Pablos}, Aitor and Cuadros, Montse and Melero, Maite},
  journal = {Proceedings of the Language Resources and Evaluation Conference},
  number  = {June},
  pages   = {3751--3760},
  title   = {{Spanish Datasets for Sensitive Entity Detection in the Legal Domain}},
  url     = {https://aclanthology.org/2022.lrec-1.400},
  year    = {2022}
}

License

CC BY 4.0