Edit model card

About

Since their beginnings in the 1830s and 1840s, news agencies have played an important role in the national and international news market, aiming to deliver news as fast and as reliable as possible. While we know that newspapers have been using agency content for a long time to produce their stories, the amount to which the agencies shape our news often remains unclear. Although researchers have already addressed this question, recently by using computational methods to assess the influence of news agencies at present, large-scale studies on the role of news agencies in the past continue to be rare.

This project aimed at bridging this gap by detecting news agencies in a large corpus of Swiss and Luxembourgish newspaper articles (the impresso corpus) for the years 1840-2000 using deep learning methods. For this, we first build and annotate a multilingual dataset with news agency mentions, which we then use to train and evaluate several BERT-based agency detection and classification models. Based on these experiments, we choose two models (for French and German) for the inference on the impresso corpus.

Model Details

The base of the model is dbmdz/bert-base-french-europeana-cased finetuned for 3 epochs on text of 256 maximum length.

Research Summary

Results show that ca. 10% of the articles explicitly reference news agencies, with the greatest share of agency content after 1940, although systematic citation of agencies already started slowly in the 1910s. Differences in the usage of agency content across time, countries and languages as well as between newspapers reveal a complex network of news flows, whose exploration provides many opportunities for future work.

Dataset Characteristics

The dataset contains 1,133 French and 397 German annotated documents, with 1,058,449 tokens, of which 1,976 have annotations. Below is an overview of the corpus statistics: The annotated dataset is released on Zenodo.

Overview of corpus statistics. %noisy gives the percentage of agency mentions with at least one OCR error.

Lg. Docs Tokens Mentions %noisy
Train de 333 247,793 493
fr 903 606,671 1,122
Total 1,236 854,464 1,615
Dev de 32 28,745 26
fr 110 77,746 114
Total 142 106,491 140
Test de 32 22,437 58
fr 120 75,057 163
Total 152 97,494 221
All de 397 298,975 577
fr 1,133 759,474 1,399
Total 1,530 1,058,449 1,976

How to use

You can use this model with Transformers pipeline for NER.

from transformers import pipeline

nlp = pipeline("newsagency-ner", model="impresso-project/ner-newsagency-bert-fr", trust_remote_code=True)
nlp("Mon nom est François et j'habite à Paris. (Reuter)")

BibTeX entry and citation info

The code is available here.

@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"
}
Downloads last month
31
Safetensors
Model size
110M params
Tensor type
F32
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.