Nel Mgenre Multilingual

🚀 Model Card for impresso-project/nel-mgenre-multilingual

The Impresso multilingual named entity linking (NEL) model is a powerful tool for accurately linking named entities in text. It's based on mGENRE (multilingual Generative ENtity REtrieval) proposed by De Cao et al, which uses a sequence - to - sequence architecture for entity disambiguation, and is built upon mBART. The model employs constrained generation to output entity names mapped to Wikidata/QIDs. It has been adapted for historical texts and fine - tuned on the HIPE - 2022 dataset, covering a wide range of historical document types and languages.

🚀 Quick Start

from transformers import AutoTokenizer, pipeline

NEL_MODEL_NAME = "impresso-project/nel-mgenre-multilingual"
nel_tokenizer = AutoTokenizer.from_pretrained(NEL_MODEL_NAME)

nel_pipeline = pipeline("generic-nel", model=NEL_MODEL_NAME,
                        tokenizer=nel_tokenizer,
                        trust_remote_code=True,
                        device='cpu')

sentence = "Le 0ctobre 1894, [START] Dreyfvs [END] est arrêté à Paris, accusé d'espionnage pour l'Allemagne — un événement qui déch1ra la société fr4nçaise pendant des années."
print(nel_pipeline(sentence))

Output Format

[
    {
        'surface': 'Dreyfvs', 
        'wkd_id': 'Q171826', 
        'wkpedia_pagename': 'Alfred Dreyfus', 
        'wkpedia_url': 'https://fr.wikipedia.org/wiki/Alfred_Dreyfus', 
        'type': 'UNK', 
        'confidence_nel': 99.98, 
        'lOffset': 24, 
        'rOffset': 33}]

The type of the entity is UNK because the model was not trained on the entity type. The confidence_nel score indicates the model's confidence in the prediction.

✨ Features

• Multilingual Support: Capable of handling over 100 languages, with optimizations for French, German, and English.
• Historical Text Adaptation: Fine - tuned on the HIPE - 2022 dataset, making it suitable for historical document analysis.
• Entity Linking: Accurately links named entities to Wikidata/QIDs using constrained generation.

📦 Installation

Since the model uses the transformers library, you can install it via pip:

pip install transformers

📚 Documentation

Model Details

Model Description

• Developed by: EPFL from the Impresso team. The project is an interdisciplinary one focused on historical media analysis across languages, time, and modalities. It is funded by the Swiss National Science Foundation (CRSII5_173719, CRSII5_213585) and the Luxembourg National Research Fund (grant No. 17498891).
• Model type: mBART - based sequence - to - sequence model with constrained beam search for named entity linking
• Languages: Multilingual (100+ languages, optimized for French, German, and English)
• License: AGPL v3+
• Finetuned from: facebook/mgenre-wiki

Model Architecture

• Architecture: mBART - based seq2seq with constrained beam search

Training Details

Training Data

The model was trained on the following datasets:

Dataset alias	README	Document type	Languages	Suitable for	Project	License
ajmc	link	classical commentaries	de, fr, en	NERC - Coarse, NERC - Fine, EL	AjMC
hipe2020	link	historical newspapers	de, fr, en	NERC - Coarse, NERC - Fine, EL	CLEF - HIPE - 2020
topres19th	link	historical newspapers	en	NERC - Coarse, EL	Living with Machines
newseye	link	historical newspapers	de, fi, fr, sv	NERC - Coarse, NERC - Fine, EL	NewsEye
sonar	link	historical newspapers	de	NERC - Coarse, EL	SoNAR

Use Cases

• Entity disambiguation in noisy OCR settings
• Linking historical names to modern Wikidata entities
• Assisting downstream event extraction and biography generation from historical archives

Limitations

• Sensitive to tokenisation and malformed spans
• Accuracy degrades on non - Wikidata entities or in highly ambiguous contexts
• Focused on historical entity mentions — performance may vary on modern texts

Environmental Impact

• Hardware: 1x A100 (80GB) for finetuning
• Training time: ~12 hours
• Estimated CO₂ Emissions: ~2.3 kg CO₂eq

📄 License

This model is licensed under AGPL v3+.

Contact

• Website: https://impresso-project.ch