Elamur Research⌕
☰
home / papers / 2026 / Comparing and Modeling Argumentation in German Political Communication across Arenas

Comparing and Modeling Argumentation in German Political Communication across Arenas

Paper · 2026 · Nina Vikhrova, Johannes Kuhling, Sebastian Haunss et al.

Summary

The paper asks whether political argumentation changes with the venue where it happens, and whether current language models can find arguments automatically. The authors built a German corpus of about 17k sentences on a single topic, COVID-19, so that differences reflect the venue rather than the subject. It covers three arenas: Bundestag plenary speeches, health committee protocols and government press conferences (Bundespressekonferenz).

Annotators marked spans where a claim, meaning a demand, plan or recommendation, is backed by a justification. Each justification received a type, a flag for whether it falls inside the subject domain, and a thematic category.

Contrary to the authors' hypothesis, press conferences argue least often but rely most on domain expertise. Plenary speeches lean mostly on normative reasoning.

A zero-shot pilot with five open-weight LLMs showed that span boundaries are hard to pin down. The models also misjudge how dense arguments are: they rarely accept that a text contains none. A baseline treating every sentence as an argument beat the best LLMs on F1 in two arenas.

The work matters as a rare comparison across arenas and as evidence about how LLMs fail at this task.

Key points

  • Data sources: GermaParl for plenary speeches, Bundestag archive pages for health committee protocols, and jungundnaiv.de transcripts for press conferences.

  • Coverage and filtering: the period runs from early 2020 to April 2023. A COVID dictionary of 1,958 keywords, based on Leibniz IDS terms, reduced 118,745 speeches, 205 protocols and 512 press conferences to 9,737 relevant texts.

  • Annotation scheme: justifications fall into three families:

    • evidence-based (disciplinary);
    • normative/pragmatic;
    • ideological/analogical.

    Each also carries an in-domain flag and a medical vs. non-medical split. Arguments are sorted into six subject classes: healthcare, containment, social, research, international and logistics.

  • Argument density: gold densities are roughly 22 arguments per 100 sentences in the Bundestag, 12 in the committee and 3 in press conferences. About 80% (406 of 490) of press-conference contributions contain no argument at all.

  • Model results:

    • All models produced too few arguments on Bundestag text.
    • Most produced too many on press conferences; Gemma3 predicted 26 per 100 sentences there.
    • Phi4mini matched the gold density best on plenary and committee data, and QWen3 came closest on press conferences.
    • The more a model's predicted density deviated from the gold density, the lower its F1 (Spearman ρ=−0.56, n=12, p=0.06).
    • The every-sentence baseline scored high F1 but near-zero or negative Gamma, which shows its predictions are essentially random.

Interesting ideas

  • I1. Checking how often a system predicts a phenomenon against how often it really occurs is a cheap diagnostic, and here it tracked overall quality fairly well.
  • I2. Pairing a trivial "label everything" baseline with a chance-corrected agreement score shows when a high F1 is just an artefact of dense data.
  • I3. Instruction-tuned models tend to accept the instruction's premise, here that arguments exist, so they hesitate to return an empty result.
  • I4. Keeping the topic fixed while varying the venue isolates register effects from vocabulary shifts.
  • I5. Labelling a claim's support as evidential, normative or rhetorical, with a separate in-domain flag, gives an interpretable layer above plain claim detection.

For Mova

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

  • I1 (learn): log predicted vs. gold rates of frames, coreference links or speech acts for each genre, and flag a run when the rates drift apart.
  • I2 (learn): add such degenerate baselines as negative controls, and report a chance-corrected agreement figure next to F1 whenever we evaluate span extraction.
  • I3 (answer): include test items whose correct reply is "nothing / not stated" so that LLM teachers and our readers are penalised for inventing content.
  • I4 (tree): assess the perceptron parser and gparse on same-topic text drawn from different registers (speech, dialogue, Q&A) to measure pure register degradation.
  • I5 (frames): attach a justification type and a domain flag to the argue/claim speech acts in prag, so that a "because" clause records what kind of support it offers.

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

🎓University of StuttgartStuttgart · Germany91 works · 46,376 citations🎓University of BremenBremen · Germany6 works · 38,395 citations

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksSocial Sciences · 1,772 worksPhilosophy · 284 worksGeneral Social Sciences · 227 works

Cite

Cite as Nina Vikhrova, Johannes Kuhling, Sebastian Haunss and Sebastian Pado. 2026. Comparing and Modeling Argumentation in German Political Communication across Arenas. arXiv:2608.00288.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2608.00288 · arXiv 2608.00288

@misc{vikhrova2026comparing,
  title = {{Comparing and Modeling Argumentation in German Political Communication across Arenas}},
  author = {Nina Vikhrova, Johannes Kuhling, Sebastian Haunss and Sebastian Pado},
  year = {2026},
  doi = {10.48550/arxiv.2608.00288},
  eprint = {2608.00288},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2608.00288}
}

References

34 references, 6 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abdelrahman Abouelenin, Atabak Ashfaq, Adam Atkinson, Hany Awadalla, Nguyen Bach, Jianmin Bao, Alon Benhaim, Martin Cai, Vishrav Chaudhary, Congcong Chen, Dong Chen, Dongdong Chen, Junkun Chen, Weizhu Chen, Yen-Chun Chen, Yi ling Chen, Qi Dai, Xiyang Dai, Ruchao Fan, and 55 others. 2025. Phi-4-Mini technical report: Compact yet powerful multimodal language models via mixture-of-LoRAs. Preprint, arXiv:2503.01743. ↗
  2. Corina Andone. 2016. Argumentative patterns in the political domain: The case of European parliamentary committees of inquiry. Argumentation, 30(1):45–60. ↗
  3. Andreas Blaette and Christoph Leonhardt. 2023. GermaParl corpus of plenary protocols. Technical report, zenodo. [Data set]. find
  4. Nico Blokker, Erenay Dayanik, Gabriella Lapesa, and Sebastian Padó. 2020. Swimming with the tide? positional claim detection across political text types. In Proceedings of the Fourth Workshop on Natural Language Processing and Computational Social Science, pages 24–34, Online. Association for Computational Linguistics. find
  5. André Bächtiger, John S. Dryzek, Jane Mansbridge, and Mark E. Warren. 2018. Deliberative Democracy. An Introduction, page 1–31. Oxford Handbooks. Oxford University Press, Oxford. find
  6. Elena Cabrio and Serena Villata. 2018. Five years of argument mining: A data-driven analysis. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI 2018), pages 5427–5433. find
  7. Johannes Daxenberger, Steffen Eger, Ivan Habernal, Christian Stab, and Iryna Gurevych. 2017. What is the essence of a claim? Cross-domain claim identification. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2055–2066, Copenhagen, Denmark. Association for Computational Linguistics. find
  8. Neele Falk, Andreas Waldis, and Iryna Gurevych. 2024. Overview of PerspectiveArg2024 the first shared task on perspective argument retrieval. In Proceedings of the 11th Workshop on Argument Mining (ArgMining 2024), pages 130–149, Bangkok, Thailand. Association for Computational Linguistics. find
  9. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 others. 2024. The Llama 3 herd of models. Preprint, arXiv:2407.21783. ↗
  10. Shohreh Haddadan, Elena Cabrio, and Serena Villata. 2019. Yes, we can! mining arguments in 50 years of US presidential campaign debates. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4684–4690, Florence, Italy. Association for Computational Linguistics. find
  11. Lore Hayek. 2024. Media framing of government crisis communication during Covid-19. Media and Communication, 12:7774. ↗
  12. Eduard Hovy, Mitchell Marcus, Martha Palmer, Lance Ramshaw, and Ralph Weischedel. 2006. OntoNotes: The 90% solution. In Proceedings of the Human Language Technology Conference of the NAACL, Companion Volume: Short Papers, pages 57–60, New York City, USA. Association for Computational Linguistics. ↗
  13. Jung & Naiv. undated. Bundespressekonferenz transcripts. Accessed 25.06.2025. find
  14. Kishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, Gaël Liu, and 196 others. 2025. Gemma 3 technical report. Preprint, arXiv:2503.19786. ↗
  15. Christer Karlsson, Thomas Persson, and Moa Mårtensson. 2024. Do members of parliament express more opposition in the plenary than in the committee? Comparing frontstage and backstage behaviour in five national parliaments. Parliamentary Affairs, 77(1):173–195. find
  16. Jan-Christoph Klie, Michael Bugert, Beto Boullosa, Richard Eckart de Castilho, and Iryna Gurevych. 2018. The inception platform: Machine-assisted and knowledge-oriented interactive annotation. In Proceedings of System Demonstrations of the 27th International Conference on Computational Linguistics (COLING 2018), Santa Fe, New Mexico, USA.
  17. Leibniz IDS. undated. Word list for the COVID-19 pandemic. Accessed 25.06.2025. find
  18. Andreas Liesenfeld, Alianda Lopez, and Mark Dingemanse. 2023. Opening up ChatGPT: Tracking openness, transparency, and accountability in instruction-tuned text generators. In Proceedings of the 5th International Conference on Conversational User Interfaces, page 1–6. Association for Computing Machinery. find
  19. Marco Lippi and Paolo Torroni. 2016. Argument mining from speech: Detecting claims in political debates. Proceedings of the AAAI Conference on Artificial Intelligence, 30(1). find
  20. Yann Mathet, Antoine Widlöcher, and Jean-Philippe Métivier. 2015. The unified and holistic method gamma (γ) for inter-annotator agreement measure and alignment. Computational Linguistics, 41(3):437–479. ↗
  21. John P. Nordin and Edward Schiappa. 2024. Argumentation: Keeping Faith with Reason, 2 edition. Routledge, New York. find
  22. Jan Pfister, Julia Wunderle, and Andreas Hotho. 2025. LLäMmlein: Transparent, compact and competitive German-only language models from scratch. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2227–2246, Vienna, Austria. Association for Computational Linguistics.
  23. Maria Poiaganova and Manfred Stede. 2025. From debates to diplomacy: Argument mining across political registers. In Proceedings of the 12th Argument mining Workshop, pages 205–216, Vienna, Austria. Association for Computational Linguistics. find
  24. Teun van Dijk. 2000. Ideology: A Multidisciplinary Approach. SAGE Publications Ltd, London. ↗
  25. J. Visser, B. Konat, R. Duthie, M. Koszowy, K. Budzynska, and C. Reed. 2020. Argumentation in the 2016 us presidential elections: annotated corpora of television debates and social media reaction. Language Resources and Evaluation, 54:123–154. ↗
  26. Henning Wachsmuth, Nona Naderi, Yufang Hou, Yonatan Bilu, Vinodkumar Prabhakaran, Tim Alberd-ingk Thijm, Graeme Hirst, and Benno Stein. 2017. Computational argumentation quality assessment in natural language. In Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers, page 176–187, Valencia, Spain. Association for Computational Linguistics.
  27. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, and 41 others. 2025. Qwen3 technical report. Preprint, arXiv:2505.09388. ↗
  28. A Ines Reinig, Ines Rehbein, and Simone Paolo Ponzetto. 2024. How to do politics with words: Investigating speech acts in parliamentary debates. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), page 8287–8300, Torino, Italia. ELRA and ICCL. find
  29. Antonio Reyes. 2011. Strategies of legitimization in political discourse: From words to actions. Discourse & Society, 22(6):781–807. ↗
  30. Ramon Ruiz-Dolz and John Lawrence. 2023. Detecting argumentative fallacies in the wild: Problems and limitations of large language models. In Proceedings of the 10th Workshop on Argument Mining, pages 1–10, Singapore. Association for Computational Linguistics. find
  31. Robin Schaefer, René Knaebel, and Manfred Stede. 2022. On selecting training corpora for cross-domain claim detection. In Proceedings of the 9th Workshop on Argument Mining, pages 181–186, Online and in Gyeongju, Republic of Korea. International Conference on Computational Linguistics. find
  32. Marco R. Steenbergen, André Bächtiger, Markus Spörndli, and Jürg Steiner. 2003. Measuring political deliberation: A discourse quality index. Comparative European Politics, 1(1):21–48. ↗
  33. Timothy Steffensmeier and William Schenck-Hamlin. 2008. Argument quality in public deliberations. Argumentation and Advocacy, 45(1):21–36. ↗
  34. Stephen E. Toulmin. 2003. The Uses of Argument, 2 edition. Cambridge University Press. Full Annotation Guidelines (follow on next complete page) Annotationsrichtlinien - Argumentation Es sollten Argumente in den gegebenen Texten annotiert werden. Ein Argument kann ein oder mehrere Sätze enthalten. Ein Argument besteht aus Begründungen und einer Behauptung, die begründet wird. Schritte zur Annotation: 1. Finde eine Behauptung im Text (→ Definition von Argumenten) 2. Kennzeichne den Text um die Behauptung (ca. 2-3 Sätze) 2. 1. Weise ein Thema dem Abschnitt zu (→ Annotation von Themen) 2.2. Kennzeichne die Begründungen, die innerhalb dieses Textabschnitts enthalten sind (→ Definition von Argumenten) 2.2.1 Gebe bei jeder Begründung an, ob es sich um eine disziplinäre Begründung handelt (→ Annotation von Argumenten) 2.2.1.1 wenn Begründung (Evidenz-theoretisch), dann weitere, binäre Unterscheidung in medizinisch oder nicht medizinisches Begründung Bemerkungen: • Fokus auf eindeutige Abschnitte • unsichere Annotation vermeiden, dennoch sollen bei impliziter Forderung oder Schlussfolgerung das Argument annotiert werden, solange die implizite Forderung erkannt wird und das Annotationsschema funktioniert, d.h. 2.1. und 2.2. können angewandt werden • Eine Schlussfolgerung in einem Argument darf eine Begründung in einem anderen sein Beispiel Meine sehr verehrten Damen und Herren, die Fraktion der FDP vertritt in ihrem Antrag die Auffassung, dass es in der Bevölkerung eine Unsicherheit über Teststrategien auf das Corona-Virus gebe und sich diese zuletzt weiter vergrößert habe. Es müsse für die Menschen nachvollziehbar sein, nach welchen Kriterien die begrenzten Test-Ressourcen verwendet würden. Daher müsse die nationale Teststrategie an den Prinzipien der Zielgerichtetheit und des Risikogruppenschutzes ausgerichtet werden. Die Strategie sowie die Quarantäne und Testempfehlungen müssten stetig anhand der wissenschaftlich aktualisierten Infektionsdynamik und Viruslast angepasst werden, da sich die wissenschaftliche Datenbasis zu SARS-CoV2 ständig erweitere. Im Einzelnen fordert die Fraktion unter anderem, Altenpflegekräfte alle 14 Tage zu testen, um Risikogruppen zu schützen, und Patienten mit Symptomen weiterhin priorisiert zu testen. Außerdem solle dafür Sorge getragen werden, dass es mit dem Ausbau digitaler Meldewege und dem 24/7-Betrieb von Laboren und Gesundheitsämtern zu deutlich weniger Verzögerungen der Ergebnisübermittlungen und eventueller Maßnahmen komme. Ich freue mich auf eine angeregte Diskussion mit unseren Sachverständigen zu diesen Fragen. Argument 1: Es müsse für die Menschen nachvollziehbar sein, nach welchen Kriterien die begrenzten Test-Ressourcen verwendet würden. Daher müsse die nationale Teststrategie an den Prinzipien der Zielgerichtetheit und des Risikogruppenschutzes ausgerichtet werden. Argument 2: Die Strategie sowie die Quarantäne und Testempfehlungen müssten stetig anhand der wissenschaftlich aktualisierten Infektionsdynamik und Viruslast angepasst werden, da sich die wissenschaftliche Datenbasis zu SARS-CoV2 ständig erweitere. Argument 3: Im Einzelnen fordert die Fraktion unter anderem, Altenpflegekräfte alle 14 Tage zu testen, um Risikogruppen zu schützen, und Patienten mit Symptomen weiterhin priorisiert zu testen. Argument 4: Außerdem solle dafür Sorge getragen werden, dass es mit dem Ausbau digitaler Meldewege und dem 24/7-Betrieb von Laboren und Gesundheitsämtern zu deutlich weniger Verzögerungen der Ergebnisübermittlungen und eventueller Maßnahmen komme. ■ – Behauptung ■ - Begründung Definition von Argumenten Argumentation: besteht aus einer Behauptung und Begründungen, wobei nicht immer beide Elemente explizit vorhanden sind. Die Begründungen sollen gekennzeichnet werden. In schriftlichen Quellen können Argumente impliziter, bspw. eher als Schlussfolgerung/Erklärung auftreten. Behauptung: find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/vikhrova-2026-comparing-modeling-argumentation-german-political/