Model Card for impresso-project/ner-newsagency-bert-multilingual

This model is designed to detect mentions of news agencies in historical newspaper articles in French and German. It was developed as part of the Impresso project, a multidisciplinary initiative aiming to enable exploration of large-scale historical media archives.

The model is fine-tuned from dbmdz/bert-base-historic-multilingual-cased, trained on a custom annotated dataset of over 1,500 historical articles (1840–2000) from the Swiss and Luxembourgish press.

Model Details

Description

• Developed by: DHLAB, EPFL and the Impresso team. The project is an interdisciplinary project focused on historical media analysis across languages, time, and modalities. Funded by the Swiss National Science Foundation (CRSII5_173719, CRSII5_213585) and the Luxembourg National Research Fund (grant No. 17498891).
• Model type: BERT-based token classification model for named entity recognition
• Languages: French and German
• License: AGPL v3+
• Finetuned from: dbmdz/bert-base-historic-multilingual-cased
• Training data: Zenodo dataset of agency mentions

Model Architecture

The model architecture consists of the following component:

• A pre-trained BERT encoder (multilingual historic BERT) as the base for token classification.

Entity Types Supported

The model predicts whether a given token span corresponds to a news agency mention. The following tags are used:

Recognized News Agencies

How to Use

from transformers import pipeline
nlp = pipeline("newsagency-ner", model="impresso-project/ner-newsagency-bert-multilingual", trust_remote_code=True)
nlp("La dépêche vient de (Reuter), diffusée hier.")

Example Output

[
  {
    'type': 'Reuters',
    'confidence': 0.99,
    'index': 5,
    'surface': 'Reuter',
    'start': 21, 'end': 27
  }
]

Training Details

Training Data

The model was trained on a custom dataset of 1,530 documents (1,133 FR / 397 DE) from the Impresso HIPE-2020 dataset. The training data was manually annotated to identify mentions of news agencies in historical newspaper articles.

Dataset Characteristics

• Annotation tool: INCEpTION
• Tags used: specific agency names, ag, pers.ind.articleauthor, unk
• Tag format: span-level BIO-style

Training Procedure

Training Hyperparameters

• Pretrained base model: dbmdz/bert-base-historic-multilingual-cased
• Finetuning epochs: 3
• Max sequence length: 512 tokens
• Languages: French and German

Evaluation

Model evaluation was performed on a manually annotated dataset with 1,530 documents (1,133 FR / 397 DE). Results are summarised below:

Evaluation methodology and full results are detailed in the thesis: Marxen, 2023. are detailed in the thesis: Marxen, 2023.

Speeds, Sizes, Times

• Model size: ~500MB
• Training time: ~half hour on 1 NVIDIA GPU (16GB)

Environmental Impact

• Hardware: 1x NVIDIA GPU (16GB)
• Training time: ~1 hour
• Provider: Local HPC (Switzerland)
• Estimated CO₂ Emissions: ~0.022 kg CO₂eq (calculated using MLCO2 calculator)

Limitations and Risks

• Does not detect news agency content without explicit mention.
• May miss OCR-degraded mentions despite robustness strategies.
• Only covers French and German.

Model Sources

• Repository: https://github.com/impresso/newsagency-classification
• Paper: Marxen, 2023
• Demo: Impresso project

Citation

@misc{marxen_newsagency_2023,
  title = {Where Did the News come from? Detection of News Agency Releases in Historical Newspapers},
  author = {Marxen, Lea and Ehrmann, Maud and Boros, Emanuela},
  year = {2023},
  url = {https://github.com/impresso/newsagency-classification/},
  note = {Master Thesis}
}

Contact

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