This repository holds the data, code, and results for a master thesis with the Hochschule für Politik at the Technical University of Munich. It includes all ground truth labels and a fully labeled version of the GC News data. To comply with copyright laws, the journalistic contents were removed from the GC News data and are available for replication upon request. The repository further includes the Python and R code used for model creation, fine-tuning, and evaluation, as well as for regression analysis and data visualization.

A labeled version of the entire corpus of German news paragraphs is included in data/gcnews/gc_news_labeled.csv. To comply with copyright laws, the journalistic contents were removed from this file. With the labels and other meta information, this file can be used to reproduce the regression and data analysis conducted in this study. A second data/gcnews/gc_news_unlabeled_sample1000.csv file contains 1000 random sample paragraphs from the GC News dataset and can be used to understand, track, and reproduce the labeling process.

The data/ground_truth/labeled_data.csv and data/ground_truth/labeled_data_unbalanced.csv files contain a stratified and a balanced version of the ground truth labels that were used to train, fine-tune, and evaluate the five implemented classification methods.

The data/external data includes the daily 7-day incidence and fatality numbers of Covid-19 in Germany, as recorded by the Robert Koch Institute. The data was initially published in

• Robert Koch-Institut (2023): 7-Tage-Inzidenz der COVID-19-Fälle in Deutschland, Berlin: Zenodo. DOI: 10.5281/zenodo.10207072
• Robert Koch-Institut (2023): COVID-19-Todesfälle in Deutschland, Berlin: Zenodo. DOI: 10.5281/zenodo.10207073

This publication applies three dictionaries via the multidictR and popdictR packages. Together with Gründl`s dictionary on populism, these were published in

• Gründl, J. (2022). Populist ideas on social media: A dictionary-based measurement of populist communication. New Media & Society, 24(6), 1481–1499. https://doi.org/10.1177/1461444820976970

The thiele_covid_terms.csv and thiele_econ_terms.csv dictionaries were originally published @

• Thiele, D. (2022). Pandemic Populism? How Covid-19 Triggered Populist Facebook User Comments in Germany and Austria. Politics and Governance, 10(1), 185–196. https://doi.org/10.17645/pag.v10i1.4712

The code in the text_classification/gzip folder implements a compressor-based k-NN-classifier and relies on Jiang et al.`s npc_gzip library. The Python codebase required for reproduction is available on pypi.org via

pip install npc-gzip

and was published in

• Jiang, Z., Yang, M., Tsirlin, M., Tang, R., Dai, Y., & Lin, J. (2023). “Low-Resource” Text Classification: A Parameter-Free Classification Method with Compressors. Findings of the Association for Computational Linguistics: ACL 2023, 6810–6828. https://aclanthology.org/2023.findings-acl.426

The exact model specifications and usage instructions for the SetFit classifier are available at https://huggingface.co/baunef/PopFit. The respective Jupyter notebook text_classification/SetFit/setfit_inference.ipynb downloads the model from HuggingFace and can be used for inference. The text_classification/SetFit/setfit_training.ipynb notebook was deployed via Google Colaboratory and trains a SetFit classifier.

The text_classification/zero_shot/zero_shot.ipynb notebook contains an inference pipeline for BERT language models and is set for testing and inference with the PopBERT classifier (https://huggingface.co/luerhard/PopBERT).

the regression and visualization directories contain R files that produce all figures and results reported in the study. The figures can be viewed in the figures directory.

If you use parts of this repository in your research, please consider citing