This repository contains the code related to the master thesis of Lea Marxen on the detection and study of news agency mentions in historical news articles from the impresso corpus (see also the thesis report Where Did the News come from?).
The project was carried out in the spring semester of 2023 under the supervision of Maud Ehrmann, Emanuela Boros and Marten Düring.
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 quickly and reliably as possible. While we know that newspapers have long used agency content to produce their stories, the extent to which the agencies shape our news is often unclear. Although researchers have already addressed this question, most recently by using computational methods to assess the influence of news agencies in today's news, large-scale studies of the role of news agencies in the past remain rare.
This master thesis aims to fill 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. We first build and annotate a two datasets (in French and German) with news agency mentions, which we then use to train and evaluate several BERT-based news agency detection and classification models. Based on these experiments, we choose two models for the inference on the whole impresso corpus.
The results show that about 10% of the articles make explicit reference to news agencies, with the largest proportion of agency content occurring after 1940, although the systematic citation of news agencies began slowly in the 1910s. Differences in the use of agency content across time, countries and languages, as well as between newspapers, reveal a complex network of news flows, the exploration of which offers many opportunities for future work.
annotation_settings/: Contains the planning and settings for annotation with Inception (DHLAB's instance).inception_settings/: Contains specifications for inception settings, e.g. the TypeSystem and tagset.Annotation Guidelines for Newsagency Classification v2.pdfannotation_planning_per_annotator.csvannotation_planning_per_doc.csv
data/: Contains part of the data used during the project.annotation/sampling/split/
lib/: Contains python scripts for the classification and the conversion of data (before and after annotation in Inception).bert-classification/: Contains the token and the text classification model as well as the code for its application on the impresso corpus.single_task: Contains on the token classification model and the code for pushing it to HuggingFace.
inception_postprocessing/inception_preprocessing/
notebooks/: Contains the notebooks used in the creation of the training corpus, annotation and analysis.1_sampling_training_data/2_annotation/3_classification/4_analysis/
report/: Contains the report of the master project (PDF and Zip of Latex sources).
The project uses python 3.10. The dependencies for running the code can be found in requirements.txt. If only the classification is required, it suffices to install the dependencies specified in lib/bert-classification/requirements_classification.txt.
The model is based on AutoModelForSequenceClassification provided by the transformers library and it is a generic model class that will be instantiated as one of the sequence classification model classes of the library when created with the from_pretrained(pretrained_model_name_or_path) method.
The classification baseline has three modes:
- binary classification: is there any news agency mentioned in the text or not
- multiclass classification: a text can only belong to a news agency
- multiclass and multilabel classification: a text can belong to several news agencies
Additional to the in-model evaluation, the HIPE-scorer can be downloaded for evaluation on the task of named entity recognition. It provides the possibility to evaluate on different time periods and OCR-levels. In order for the latter to work, we changed row["MISC"] to row["OCR-INFO"] in line 200 of HIPE-scorer/hipe_evaluation/utils.py.
The two fine-tuned models used for inference on the impresso corpus are released on Zenodo.
The annotated dataset is released on Zenodo. It contains historical newspaper articles with annotations of news agency mentions. The articles are divided into French (fr) and German (de) subsets and a train, dev and test set respectively. The data is annotated at token level in the CoNLL format with IOB tagging format.
| Lg. | Docs | Tokens | Agency Mentions | |
|---|---|---|---|---|
| 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 |
Due to an error, there are seven duplicated articles in the French test set (article IDs: courriergdl-1847-10-02-a-i0002, courriergdl-1852-02-14-a-i0002, courriergdl-1860-10-31-a-i0016, courriergdl-1864-12-15-a-i0005, lunion-1860-11-27-a-i0004, lunion-1865-02-05-a-i0012, lunion-1866-02-16-a-i0009).
The dataset contains the following news agencies:
# global.columns = TOKEN NE-COARSE-LIT NE-COARSE-METO NE-FINE-LIT NE-FINE-METO NE-FINE-COMP NE-NESTED NEL-LIT NEL-METO RENDER SEG OCR-INFO MISC
# language = fr
# newspaper = EXP
# date = 1924-03-27
# document_id = EXP-1924-03-27-a-i0077
# news-agency-as-source = Q2826560
# segment_iiif_link = https://impresso-project.ch/api/proxy/iiif/EXP-1924-03-27-a-p0005/224,107,285,87/full/0/default.jpg
POLITIQUE O O O O O O _ _ EndOfLine _ _ _
# segment_iiif_link = https://impresso-project.ch/api/proxy/iiif/EXP-1924-03-27-a-p0005/129,269,488,53/full/0/default.jpg
LONDRES O O O O O O _ _ NoSpaceAfter _ _ _
, O O O O O O _ _ _ _ _ _
27 O O O O O O _ _ _ _ _ _
( O O O O O O _ _ NoSpaceAfter _ _ _
Havaa B-org O B-org.ent.pressagency.Havas O O O Q2826560 _ NoSpaceAfter _ Transcript:Havas|LED0.20 _
) O O O O O O _ _ NoSpaceAfter _ _ _
. O O O O O O _ _ _ EndOfSentence _ _
...
This article contains the news agency Havas and thus its WikiID Q2826560 appears on article-level (# news-agency-as-source).
- Copyright (c) 2023 for the code: Lea Marxen and EPFL
- Code license: MIT
This thesis was carried out in the context of the 'impresso - Media Monitoring of the Past' projects, which aim to develop new approaches to the exploration of historical media content across time, language and national borders.
The first impresso project (2017-2020) compiled and semantically enriched a corpus of digitised Swiss and Luxembourg newspapers and designed a system architecture and user interface, which together form the impresso app. It was supported by the Swiss National Science Foundation (grant No. CRSII5_173719).
The second, follow-up project (2023-2027) 'impresso - Media Monitoring of the Past II. Beyond Borders: Connecting Historical Newspapers and Radio' continues this effort and proposes to overcome language and media barriers and, for the first time, to enable the joint exploration of newspaper and radio archive contents across time, language and national borders. Leveraging an unprecedented corpus of transnational print and broadcast media, it aims to enrich and connect these sources into a common vector space, and to design appropriate, meaningful and transparent exploration capabilities for historical research from a transmedia and transnational perspective. It is supported by the Swiss National Science Foundation (grant No. CRSII5_213585) and the Luxembourg National Research Fund (grant No. 17498891).