Dataset and Models for Detection of News Agency Releases in Historical Newspapers

<p>This record contains the annotated datasets and models used and produced for the work reported in the Master Thesis &quot;<em>Where Did the News come from? Detection of News Agency Releases in Historical Newspapers</em> &quot; (<a href="https://infoscience.epfl.ch/record/305129?&amp;ln=en">link</a>).</p> <p>Please cite this report if you are using the models/datasets or find it relevant to your research:</p> <pre><code>@article{Marxen:305129, title = {Where Did the News Come From? Detection of News Agency Releases in Historical Newspapers}, author = {Marxen, Lea}, pages = {114p}, year = {2023}, url = {http://infoscience.epfl.ch/record/305129}, }</code></pre> <p><br> <strong>1. DATA</strong></p> <p>The <strong>newsagency-dataset</strong> 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.</p> <p>The distribution of articles in the different sets is as follows:</p> <table> <caption>Dataset Statistics</caption> <thead> <tr> <th scope="row">&nbsp;</th> <th scope="col">Lg.</th> <th scope="col">Docs</th> <th scope="col">Agency Mentions</th> </tr> </thead> <tbody> <tr> <th scope="row">Train</th> <td>de</td> <td>333</td> <td>493</td> </tr> <tr> <th scope="row">&nbsp;</th> <td>fr</td> <td>903</td> <td>1,122</td> </tr> <tr> <th scope="row">Dev</th> <td>de</td> <td>32</td> <td>26</td> </tr> <tr> <th scope="row">&nbsp;</th> <td>fr</td> <td>110</td> <td>114</td> </tr> <tr> <th scope="row">Test</th> <td>de</td> <td>32</td> <td>58</td> </tr> <tr> <th scope="row">&nbsp;</th> <td>fr</td> <td>120</td> <td>163</td> </tr> </tbody> </table> <p>Due to an error, there are seven duplicated articles in the French test set (<em>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</em>).</p> <p>&nbsp;</p> <p><strong>2. MODELS</strong></p> <p>The two agency detection and classification models used for the inference on the <em><a href="https://impresso-project.ch/">impresso</a> </em>Corpus are released as well:</p> <ul> <li><strong>newsagency-model-de</strong>: based on <a href="https://www.deepset.ai/german-bert">German BERT</a> (with maximum sequence length 128), fine-tuned with the German training set of the newsagency-dataset</li> <li><strong>newsagency-model-fr</strong>: based on <a href="https://huggingface.co/dbmdz/bert-base-french-europeana-cased">French Europeana BERT</a> (with maximum sequence length 128), fine-tuned with the French training set of the newsagency-dataset</li> </ul> <p>The models perform multitask classification with two prediction heads, one for token-level agency entity classification and one for sentence-level (<em>has_agency: yes/no</em>). They can be run with TorchServe, for details see the <a href="https://github.com/impresso/newsagency-classification/tree/main/lib/bert_classification">newsagency-classification</a> repository.</p> <p>&nbsp;</p> <p>Please refer to the report for further information or contact us.</p> <p>&nbsp;</p> <p><strong>3. CODE</strong></p> <p><a href="https://github.com/impresso/newsagency-classification">https://github.com/impresso/newsagency-classification</a></p> <p>&nbsp;</p> <p><strong>4. CONTACT</strong></p> <p>Maud Ehrmann (EPFL-DHLAB)<br> Emanuela Boros (EPFL-DHLAB)</p>