MAPA Irish NER Model (Legal Domain)

This model is part of the MAPA (Multilingual Anonymisation for Public Administrations) toolkit, developed by Pangeanic and funded by the European Union through the Connecting Europe Facility (CEF) programme.

It performs hierarchical Named Entity Recognition (NER) for the detection of personally identifiable information (PII) in Irish legal text, with the goal of supporting anonymisation workflows in public administrations.

⚠️ Note on training data: Like the German MAPA model, this model was trained on machine-translated legal data rather than on native Irish legal corpora. See the Limitations and Considerations section below.

Model Description

Developed by: Pangeanic, as part of the MAPA EU Project
Base model: google-bert/bert-base-multilingual-cased (mBERT cased)
Architecture: EnhancedTwoFlatLevelsSequenceLabellingModel — a custom BERT-based architecture with two parallel classification heads
Language: Irish (ga)
Domain: Legal
Training data: Machine-translated Irish legal documents
License: Apache 2.0

Hierarchical NER

The model performs token classification at two levels simultaneously:

Level 1 (coarse-grained): 19 entity categories (e.g. PERSON, ORGANISATION, LOCATION, DATE, ADDRESS...).
Level 2 (fine-grained): 119 entity subcategories (e.g. for a PERSON: title, given name, family name...).

Example output structure:

{
  "annotations": [
    { "content": "Uas. Connelly", "value": "PERSON" },
    { "content": "Uas.",          "value": "title" },
    { "content": "Connelly",      "value": "family name" }
  ]
}

The full label inventories are included in this repository as level1_tags_vocabulary.json and level2_tags_vocabulary.json.

Intended Use

This model is intended to be used as part of the MAPA toolkit, which provides the full anonymisation pipeline (entity detection + entity replacement).

The model uses a custom architecture (EnhancedTwoFlatLevelsSequenceLabellingModel) defined in the MAPA codebase. It is not directly loadable via AutoModel.from_pretrained from the transformers library without that code.

To use this model, clone the MAPA toolkit:

🔗 Repository: https://github.com/PangeanicAI/MAPA-EU-Project

The repository contains the model class definition, inference scripts, and the complete anonymisation pipeline (including the entity replacement module, which uses auxiliary resources not included in this HF repo).

Training Details

Training domain: Machine-translated Irish legal documents
Epochs: 60
Iterations: 9,028
Base architecture: BERT-base multilingual cased
Hidden size: 768
Layers: 12
Attention heads: 12
Vocabulary size: 119,547

Evaluation Metrics

Final metrics reported at the end of training:

Metric	Value
Level 1 micro-F1	0.8258
Level 1 binary-F1	0.8844
Level 2 micro-F1	0.8084
Final loss	0.1481

Limitations and Considerations

Machine-translated training data: This model's training corpus consists of machine-translated legal text rather than native Irish legal documents. Performance on authentic Irish legal writing — with native syntactic patterns (initial mutations, VSO order), Gaeilge-specific legal terminology and case-specific phrasings — may differ substantially from the metrics reported above, which were computed on data of the same machine-translated distribution.
Low-resource language considerations: Irish (Gaeilge) is among the most underrepresented languages in the pretraining data of mBERT. Combined with the fact that fine-tuning was performed on translated rather than native text, the resulting representations should be treated with extra caution in production settings.
Not a standalone HuggingFace model: the custom architecture requires the MAPA toolkit code to be instantiated and used. The inference widget on this page is intentionally disabled.
Domain bias: the model was trained on legal documents and may underperform on text from different domains (e.g. clinical notes, administrative communication, social media) or on Irish variants underrepresented in the (translated) training data.
Anonymisation is not guaranteed: as with any NER-based system, false negatives are possible. Outputs should be reviewed before publishing or sharing sensitive content.
No additional safety testing has been performed in this HuggingFace release. The weights are published as they were produced during the original MAPA project.

Related Models

Other models from the MAPA project are available under the Pangeanic organisation, covering additional languages and domains.

Acknowledgements

The MAPA project was funded by the European Union under the Connecting Europe Facility (CEF) programme, grant agreement INEA/CEF/ICT/A2019/1927065.

Citation

If you use this model, please cite the MAPA project:

@inproceedings{mapa2022,
  title     = {{MAPA} Project: Ready-to-Go Open-Source Datasets and Deep Learning Technology to Remove Identifying Information from Text Documents},
  author    = {Gianola, Lucia and Ajausks, \=Emils and Arranz, Victoria and Bendi, Chomicha and Choukri, Khalid and Ciulla, Montse and Coheur, Luísa and Costa, Costanza and Cruz, Elena and Esplà-Gomis, Miquel and Garcia-Martinez, Mercedes and Herranz, Manuel and Iranzo-Sánchez, Javier and Klūga, Mārcis and Labaka, Gorka and Lagzdiņš, Artūrs and Lazar, Alina and Mahdi, Mohammed and Otero, Carla and Pinnis, Mārcis and Rigau, German and Ryšavá, Klára and Saint-Dizier, Patrick and Sosoni, Vilelmini},
  booktitle = {Proceedings of the Workshop on Challenges in the Management of Large Corpora (CMLC-10)},
  year      = {2022}
}

Pangeanic/mapa-ga-legal