Elamur Research⌕
☰
home / papers / 2026 / Ideology Prediction of German Political Texts

Ideology Prediction of German Political Texts

Paper · 2026 · Sinclair Schneider, Florian Steuber, Joao A. G. Schneider et al.

Summary

The paper looks at how to place German political text on a continuous left–right axis, d ∈ [−1, 1]. Earlier work mostly sorts text into fixed classes such as left, centre and right. A continuous score lets analysts pick out any band of the spectrum. The authors built four corpora:

  • Bundestag plenary speeches, auto-labelled from recorded interjections and applause
  • statements from the Wahl-O-Mat voting-advice tool
  • articles from 33 German newspapers with known leanings
  • 535,200 tweets from 597 members of parliament

The first two corpora were used for training and the last two only for out-of-domain testing. They compared 13 transformers (BERT-family, Llama and Gemma). Each model predicts which parties would agree with a text. Every party is assigned a unit vector at its angle on the spectrum, and the angle of the weighted vector sum becomes the score. DeBERTa-large performed best in-domain and on tweets, while Gemma2-2B was best on newspapers. The authors conclude that architecture and domain-matched data can matter as much as model size for this task.

Key points

  • The Bundestag protocols cover October 2017 to September 2024 (34,174 speeches) and yield 32,246 party pro/contra statements. Labels come from interruptions such as applause or heckling, typed by party; speeches with no interruptions were dropped.
  • Llama 3.1 rewrote each text in five registers to add linguistic variety: child, teenager, adult, eloquent speaker, and tweet.
  • In-domain, DeBERTa-large reached F1 = 0.844.
  • On tweets, DeBERTa-large predicted the author's party with ACC = 0.864 once 100+ words were available.
  • On the 33 newspapers, Gemma2-2B reached MAE = 0.172, about 8.58% error against a survey-based benchmark of outlet leanings.
  • Stated limitations:
    • The scale is bounded by the outermost parties (Die Linke, AfD), so more extreme positions cannot be told apart from them.
    • AfD and Die Linke sometimes agree, possibly because both are in opposition.
    • The classifier offers no explanation of its decisions.
    • It needs retraining as political positions drift.

Interesting ideas

  • I1. Parliament transcripts record who applauded or heckled whom, which gives free stance labels without manual annotation; structural side-signals in a document can serve as a cheap weak supervisor.
  • I2. Asking an LLM to restate each training text in several registers (childlike, eloquent, tweet-like) widens stylistic coverage, so a model trained on formal speech transfers better to informal text.
  • I3. A continuous score is obtained from a categorical model by giving each class a unit vector at a meaningful angle and reading off the angle of the probability-weighted sum.
  • I4. The authors train on two sources and evaluate only on two unrelated sources, so the test measures transfer instead of memorised in-domain cues.
  • I5. Results are reported as a function of how much text is available, and accuracy rises once 100+ words are present.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): We could mine similar by-product signals (dialogue replies, quoted reactions, markup) as extra silver evidence next to the witness parsers, scoring them against negative controls.
  • I2 (tree): Register-shifted rewrites of treebank-style sentences, parsed and checked for agreement, could harden gparse and the perceptron parser against casual or social-media English.
  • I3 (world): The same trick could turn verb-class or sentiment distributions into graded values on a scale (e.g., degree of approval), while keeping the per-class contributions visible for explanation.
  • I4 (learn): Holding out entire genres, such as fables or maths problems, as test-only material would give a more honest read on whether induced repair rules generalise.
  • I5 (answer): Breaking down reader accuracy by passage length would show where answers become unreliable and when Mova should decline or hedge.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

Also on the byline: Universität der Bundeswehr München · Bundeswehr.

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 worksSocial Sciences · 1,954 worksGeneral Social Sciences · 254 works

Cite

Cite as Sinclair Schneider, Florian Steuber, Joao A. G. Schneider and Gabi Dreo Rodosek. 2026. Ideology Prediction of German Political Texts. arXiv:2605.14352.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2605.14352 · arXiv 2605.14352

@misc{schneider2026ideology,
  title = {{Ideology Prediction of German Political Texts}},
  author = {Sinclair Schneider, Florian Steuber, Joao A. G. Schneider and Gabi Dreo Rodosek},
  year = {2026},
  doi = {10.48550/arxiv.2605.14352},
  eprint = {2605.14352},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2605.14352}
}

References

50 references, 14 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Aksenov, D.; Bourgonje, P.; Zaczynska, K.; Ostendorff, M.; Moreno-Schneider, J.; and Rehm, G. 2021. Fine-grained Classification of Political Bias in German News: A Data Set and Initial Experiments. In Mostafazadeh Davani, A.; Kiela, D.; Lambert, M.; Vidgen, B.; Prabhakaran, V.; and Waseem, Z., eds., Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH), 121–131. Stroudsburg, Penn.: Association for Computational Linguistics. find
  2. AllSides. 2025. AllSides: Unbiased Balanced News. https://www.allsides.com/. Accessed 2025-09-10. ↗
  3. Andrzejewski, C. 2023. Team Jorge: In the heart of a global disinformation machine. https://forbiddenstories.org/team-jorge-disinformation. Accessed 2026-01-15. ↗
  4. Baly, R.; Da San Martino, G.; Glass, J.; and Nakov, P. 2020. We Can Detect Your Bias: Predicting the Political Ideology of News Articles. In Webber, B.; Cohn, T.; He, Y.; and Liu, Y., eds., Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 4982– 4991. Online: Association for Computational Linguistics.
  5. Baly, R.; Karadzhov, G.; Saleh, A.; Glass, J.; and Nakov, P. 2019. Multi-Task Ordinal Regression for Jointly Predicting the Trustworthiness and the Leading Political Ideology of News Media. In Burstein, J.; Doran, C.; and Solorio, T., eds., Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, volume 1, 2109–2116. Minneapolis, Minn.: Association for Computational Linguistics. ↗
  6. Boizard, N.; Gisserot-Boukhlef, H.; Alves, D. M.; Martins, A.; Hammal, A.; Corro, C.; Hudelot, C.; Malherbe, E.; Malaboeuf, E.; Jourdan, F.; Hautreux, G.; Alves, J.; Haddad, K. E.; Faysse, M.; Peyrard, M.; Guerreiro, N. M.; Fernandes, P.; Rei, R.; and Colombo, P. 2025. EuroBERT: Scaling Multilingual Encoders for European Languages. arXiv:2503.05500. ↗
  7. Bolte, F. 2025. qual-o-mat-data. https://github.com/ gockelhahn/qual-o-mat-data. Accessed 2025-09-10. ↗
  8. Chan, B.; Schweter, S.; and Möller, T. 2020. German’s Next Language Model. In Proceedings of the 28th International Conference on Computational Linguistics, 6788– 6796. Barcelona, Spain: International Committee on Computational Linguistics.
  9. Clark, K.; Luong, M.-T.; Le, Q. V.; and Manning, C. D. 2020. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. arXiv:2003.10555. ↗
  10. Cohen, R.; and Ruths, D. 2013. Classifying Political Orientation on Twitter: It’s Not Easy! In Proceedings of the International AAAI Conference on Weblogs and Social Media (ICWSM), volume 7, 91–99. Cambridge, Mass. find
  11. Conneau, A.; Khandelwal, K.; Goyal, N.; Chaudhary, V.; Wenzek, G.; Guzmán, F.; Grave, E.; Ott, M.; Zettlemoyer, L.; and Stoyanov, V. 2020. Unsupervised Cross-lingual Representation Learning at Scale. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 8440–8451. Online: Association for Computational Linguistics.
  12. Cumming, G. 2013. Understanding the new statistics: Effect sizes, confidence intervals, and meta-analysis. Routledge. find
  13. Dada, A.; Chen, A.; Peng, C.; Smith, K.; Idrissi-Yaghir, A.; Seibold, C.; Li, J.; Heiliger, L.; Friedrich, C.; Truhn, D.; Egger, J.; Bian, J.; Kleesiek, J.; and Wu, Y. 2023. On the Impact of Cross-Domain Data on German Language Models. In Findings of the Association for Computational Linguistics: EMNLP 2023, 13801–13813. Singapore: Association for Computational Linguistics. ↗
  14. Deutscher Bundestag. 2025. Open Data. https://www.bundestag.de/services/opendata. Accessed 2025-09-10. ↗
  15. Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, volume 1, 4171–4186. Minneapolis, Minn.: Association for Computational Linguistics.
  16. Dubey, A.; Jauhri, A.; and Pandey, A. 2024. The Llama 3 Herd of Models. arXiv:2407.21783. ↗
  17. Erhard, L.; Hanke, S.; Remer, U.; Falenska, A.; and Heiberger, R. H. 2025. PopBERT. Detecting Populism and Its Host Ideologies in the German Bundestag. Political Analysis, 33(1): 1–17. find
  18. Fagni, T.; and Cresci, S. 2022. Fine-grained Prediction of Political Leaning on Social Media with Unsupervised Deep Learning. Journal of Artificial Intelligence Research, 73: 633–672. FORCE11. 2020. The FAIR Data principles. https://force11.org/info/the-fair-data-principles/. ↗
  19. Gebru, T.; Morgenstern, J.; Vecchione, B.; Vaughan, J. W.; Wallach, H.; Iii, H. D.; and Crawford, K. 2021. Datasheets for datasets. Communications of the ACM, 64(12): 86–92. ↗
  20. Gemini Team Google. 2025. Gemini: A Family of Highly Capable Multimodal Models. arXiv:2312.11805. ↗
  21. Grootendorst, M. 2022. BERTopic: Neural topic modeling with a class-based TF-IDF procedure. arXiv:2203.05794. ↗
  22. He, P.; Liu, X.; Gao, J.; and Chen, W. 2020. De-BERTa: Decoding-enhanced bert with disentangled attention. arXiv:2006.03654. ↗
  23. Honeywell, C. 2021. Anarchism. Key Concepts in Political Theory. Cambridge, United Kingdom: Polity Press. find
  24. Jakob, C.; Wenzel, P.; Mohtaj, S.; and Schmitt, V. 2024. Augmented Political Leaning Detection: Leveraging Parliamentary Speeches for Classifying News Articles. In Klamm, C.; Lapesa, G.; Ponzetto, S. P.; Rehbein, I.; and Sen, I., eds., Proceedings of the 4th Workshop on Computational Linguistics for the Political and Social Sciences: Long and short papers, 126–133. Vienna, Austria: Association for Computational Linguistics. find
  25. Jiang, J.; Ren, X.; and Ferrara, E. 2023. Retweet-BERT: Political Leaning Detection Using Language Features and Information Diffusion on Social Networks. Proceedings of the International AAAI Conference on Web and Social Media, 17(1): 459–469. find
  26. Kawintiranon, K.; and Singh, L. 2022. PoliBERTweet: A Pre-trained Language Model for Analyzing Political Content on Twitter. In Calzolari, N.; Béchet, F.; Blache, P.; Choukri, K.; Cieri, C.; Declerck, T.; Goggi, S.; Isahara, H.; Maegaard, B.; Mariani, J.; Mazo, H.; Odijk, J.; and Piperidis, S., eds., Proceedings of the Thirteenth Language Resources and Evaluation Conference, 7360–7367. Marseille, France: European Language Resources Association. find
  27. Kiesel, J.; Mestre, M.; Shukla, R.; Vincent, E.; Adineh, P.; Corney, D.; Stein, B.; and Potthast, M. 2019. SemEval-2019 Task 4: Hyperpartisan News Detection. In May, J.; Shutova, E.; Herbelot, A.; Zhu, X.; Apidianaki, M.; and Mohammad, S. M., eds., Proceedings of the 13th International Workshop on Semantic Evaluation, 829–839. Minneapolis, Minn.: Association for Computational Linguistics.
  28. Krier, F. 2009. Sozialismus für Kleinbürger: Pierre Joseph Proudhon - Wegbereiter des Dritten Reiches. Köln: Böhlau. find
  29. Levy, C.; and Adams, M. S., eds. 2019. The Palgrave Handbook of Anarchism. Cham: Springer. find
  30. Liu, Y.; Zhang, X. F.; Wegsman, D.; Beauchamp, N.; and Wang, L. 2022. POLITICS: Pretraining with Same-story Article Comparison for Ideology Prediction and Stance Detection. In Carpuat, M.; de Marneffe, M.-C.; and Meza Ruiz, I. V., eds., Findings of the Association for Computational Linguistics: NAACL 2022, 1354–1374. Seattle, Wash.: Association for Computational Linguistics.
  31. Maurer, M.; Kruschinski, S.; and Jost, P. 2024. Fehlt da was? Perspektivenvielfalt in den öffentlich-rechtlichen Nachricht-enformaten. Technical report, Johannes Gutenberg-Universität Mainz, Institut für Publizistik, Mainz. find
  32. Mediabiasfactcheck.com. 2025. Media Bias/Fact Check - Search and Learn the Bias of News Media. https://mediabiasfactcheck.com/filtered-search/?country=DE. Accessed: 2026-01-04. ↗
  33. Medienkompass.org. 2025. Deutsche Medienlandschaft. https://medienkompass.org/deutsche-medienlandschaft/. Accessed: 2025-07-07. ↗
  34. Nguyen, D. Q.; Vu, T.; and Tuan Nguyen, A. 2020. BERTweet: A pre-trained language model for English Tweets. In Liu, Q.; and Schlangen, D., eds., Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations, 9–14. Online: Association for Computational Linguistics.
  35. Ortiz Suárez, P. J.; Sagot, B.; and Romary, L. 2019. Asynchronous pipelines for processing huge corpora on medium to low resource infrastructures. In Bański, P.; Barbaresi, A.; Biber, H.; Breiteneder, E.; Clematide, S.; Kupietz, M.; Lüngen, H.; and Iliadi, C., eds., Proceedings of the Workshop on Challenges in the Management of Large Corpora, 9–16. Cardiff, United Kingdom: Leibniz-Institut. ↗
  36. Ostendorff, M.; Blume, T.; and Ostendorff, S. 2020. Towards an Open Platform for Legal Information. In Proceedings of the ACM/IEEE Joint Conference on Digital Libraries in 2020, 385–388. Online: Association for Computing Machinery. find
  37. Preoţiuc-Pietro, D.; Liu, Y.; Hopkins, D.; and Ungar, L. 2017. Beyond Binary Labels: Political Ideology Prediction of Twitter Users. In Barzilay, R.; and Kan, M.-Y., eds., Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, volume 1, 729–740. Vancouver, Canada: Association for Computational Linguistics. find
  38. Proudhon, P.-J. 1840. Qu’est-ce que la propriété?, ou, Recherches sur le principe du droit et du gouvernement: Premier mémoire. Brocard. find
  39. Riviere, M.; Pathak, S.; Sessa, P. G.; Hardin, C.; and Bhupatiraju, S. 2024. Gemma 2: Improving Open Language Models at a Practical Size. arXiv:2408.00118. ↗
  40. Schapiro, J. S. 1945. Pierre Joseph Proudhon, Harbinger of Fascism. The American Historical Review, 50(4): 714. ↗
  41. Scheible, R.; Frei, J.; Thomczyk, F.; He, H.; Tippmann, P.; Knaus, J.; Jaravine, V.; Kramer, F.; and Boeker, M. 2024. find
  42. GottBERT: a pure German Language Model. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 21237–21250. Miami, Fla.: Association for Computational Linguistics. find
  43. Schneider, S. 2025a. german ideology prediction. https://github.com/SinclairSchneider/german ideology prediction. Accessed 2025-09-15. ↗
  44. Schneider, S. 2025b. trainset political party big. https://doi.org/10/qvxb. Revision 444f2da, Accessed 2025-09-10. ↗
  45. Schneider, S. 2025c. trainset political text yes no german. https://doi.org/10/qvw9. Accessed 2025-09-12. ↗
  46. Springer. 2025. BILD.de. https://www.bild.de. Accessed 2025-12-22. ↗
  47. Tiedemann, J. 2012. Parallel Data, Tools and Interfaces in OPUS. In Calzolari, N.; Choukri, K.; Declerck, T.; Doğan, M. U.; Maegaard, B.; Mariani, J.; Moreno, A.; Odijk, J.; and Piperidis, S., eds., Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12), 2214–2218. Istanbul, Turkey: European Language Resources Association (ELRA). find
  48. Volf, M.; and Simko, J. 2025. Political Leaning and Politicalness Classification of Texts. arXiv:2507.13913. ↗
  49. Zhang, Y.; Li, M.; Long, D.; Zhang, X.; Lin, H.; Yang, B.; Xie, P.; Yang, A.; Liu, D.; Lin, J.; Huang, F.; and Zhou, J. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models. arXiv:2506.05176. ↗
  50. Zhuang, L.; Wayne, L.; Ya, S.; and Jun, Z. 2021. A Robustly Optimized BERT Pre-training Approach with Post-training. In Li, S.; Sun, M.; Liu, Y.; Wu, H.; Liu, K.; Che, W.; He, S.; and Rao, G., eds., Proceedings of the 20th Chinese National Conference on Computational Linguistics, 1218– 1227. Huhhot, China: Chinese Information Processing Society of China. Paper Checklist

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/schneider-2026-ideology-prediction-german-political-texts/