Elamur Research⌕
☰
home / papers / 2025 / Analyzing German Parliamentary Speeches: A Machine Learning Approach for Topic and Sentiment Classification

Analyzing German Parliamentary Speeches: A Machine Learning Approach for Topic and Sentiment Classification

Paper · 2025 · Lukas Patz, Moritz Beyer, Jannik Spath et al.

Summary

The authors ask how topics and tone in the German Bundestag changed over roughly five years (2019–2024), framing this against declining public trust in politics. Bundestag debates are an unfiltered record of political discourse, unlike media coverage. They collected about 28,000 plenary speeches and hand-annotated a small subset for topic and sentiment. On that subset they trained two separate supervised classifiers, using custom embeddings and features designed for each task. The topic model reached an AUROC of 0.94, averaged over topics, and the sentiment model reached 0.89. They then labeled the whole corpus automatically and compared parties, time periods, and government versus opposition roles. Opposition parties are consistently more negative, and a party's style shifts once it leaves government. Parties also tend to be most critical inside their own signature policy areas. The code and data are public. The work shows that a modest amount of careful in-domain annotation can support large-scale analysis of political discourse.

Key points

  • The corpus has about 28,000 Bundestag speeches from 2019–2024. The 20th Bundestag had 733 members and holds about 80 plenary sessions a year.
  • The approach has two steps: classifiers are trained on a manually labeled subset and then applied to the full corpus for trend analysis.
  • Results are AUROC 0.94 for topic classification (mean across topics) and 0.89 for sentiment.
  • Three topic areas take up most of the debate: Foreign and Security Policy, Economy and Finance, and Social and Education. The SPD leans toward Social and Education, and Bündnis 90/Die Grünen toward Environment. CDU/CSU and FDP are more negative on Economy and Finance.
  • Limitations:
    • Themes overlap, which suggests multi-label classification is needed.
    • Annotation is subjective.
    • The labeled set is small.
  • Future work proposed: active learning, crowdsourcing, or semi-supervised labeling. Related resources include ParlSpeech V2, GerParCor, and the Richter et al. dataset of about 900k speeches (1949–2022).

Interesting ideas

  • I1. The authors built separate, task-tailored classifiers for topic and for tone instead of one generic model, arguing that each dimension needs its own features. This keeps each component simpler and easier to audit.
  • I2. A party switching from government to opposition acts as a natural contrast: the same actor appears in a different institutional role, so the change in tone can be attributed to the role. Such within-actor comparisons are cleaner than comparing different groups.
  • I3. The paper warns that negative tone on a topic may reflect general discontent rather than criticism aimed at that topic, so the target of negativity must be separated from the topic of the text.
  • I4. Misclassifications came from speeches spanning several themes, which points toward multi-label topic assignment instead of forcing a single class.
  • I5. The authors propose growing a small hand-labeled seed set through active learning or semi-supervised labeling, to cover rare classes cheaply.

For Mova

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

  • I1 (frames): Speech-act or stance detection could be kept as its own explainable module beside role labeling, so each has dedicated cues and is not merged into one opaque score.
  • I2 (learn): Evaluation sets could use paired cases where the same entity or wording appears under a changed context, testing whether the system tracks the context instead of the surface form.
  • I3 (frames): Sentiment or irony could be attached to the specific frame argument it targets, never to the whole document, which also helps flag sarcastic praise.
  • I4 (world): A text's world model could hold several active domains with weights, so later reasoning is not locked into one wrongly chosen topic.
  • I5 (learn): Silver data from agreeing witness parsers could be prioritized by disagreement, sending only the contested sentences to the LLM teacher tier or a human.

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

 Fields, topics and concepts

📚 Shelves 6The 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 worksGeneral Social Sciences · 227 works

🧭 Topics 1Research subjects and languages the work is about.Sentiment & emotion · 972 works

Cite

Cite as Lukas Patz, Moritz Beyer, Jannik Spath, Lasse Bohlen, Patrick Zschech, Mathias Kraus and Julian Rosenberger. 2025. Analyzing German Parliamentary Speeches: A Machine Learning Approach for Topic and Sentiment Classification. arXiv:2508.03181.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2508.03181 · arXiv 2508.03181

@misc{patz2025analyzing,
  title = {{Analyzing German Parliamentary Speeches: A Machine Learning Approach for Topic and Sentiment Classification}},
  author = {Lukas Patz, Moritz Beyer, Jannik Spath, Lasse Bohlen, Patrick Zschech, Mathias Kraus and Julian Rosenberger},
  year = {2025},
  doi = {10.48550/arxiv.2508.03181},
  eprint = {2508.03181},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2508.03181}
}

References

50 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Abercrombie, G. & Batista-Navarro, R. (2020), ‘Sentiment and position-taking analysis of parliamentary debates: a systematic literature review’, Journal of Computational Social Science 3(1), 245–270. ↗
  2. Abrami, G., Bagci, M., Hammerla, L. & Mehler, A. (2022), German Parliamentary Corpus (GerParCor), in ‘Proceedings of the Thirteenth Language Resources and Evaluation Conference’, Marseille, France, pp. 1900–1906. find
  3. Akter, S., Bhattacharyya, M., Wamba, S. F. & Aditya, S. (2016), ‘How does social media analytics create value?’, Journal of Organizational and End User Computing (JOEUC) 28(3), 1–9. ↗
  4. Barbaresi, A. (2018), A corpus of german political speeches from the 21st century, in ‘11th Language Resources and Evaluation Conference (LREC 2018)’, European Language Resources Association (ELRA), pp. 792–797. find
  5. Blei, D. M., Ng, A. Y. & Jordan, M. I. (2003), ‘Latent dirichlet allocation’, Journal of machine Learning research 3(Jan), 993–1022. ↗
  6. Boulianne, S. (2020), ‘Twenty years of digital media effects on civic and political participation’, Communication research 47(7), 947–966. find
  7. Breiman, L. (2001), ‘Random forests’, Machine learning 45, 5–32. find
  8. Bundesverfassungsgericht (2023), ‘Beschluss des Zweiten Senats vom 15. November 2023 - 2 BvF 1/22’. BVerfG, Beschluss vom 15. November 2023 - 2 BvF 1/22. ECLI:DE:BVerfG:2023:fs20231115.2bvf000122. Veröffentlicht u.a. in: NJW 2024, 501. find
  9. Chapman, C. & Stolee, K. T. (2016), Exploring regular expression usage and context in python, in ‘Proceedings of the 25th International Symposium on Software Testing and Analysis’, ISSTA 2016, Association for Computing Machinery, pp. 282–293. find
  10. Chen, T. & Guestrin, C. (2016), Xgboost: A scalable tree boosting system, in ‘Proceedings of the 22nd acm sigkdd international conference on knowledge discovery and data mining’, pp. 785–794. find
  11. Connolly, K. (2023), ‘Fear and anger rise in germany over handling of budget crisis’. URL: https://www.theguardian.com/world/2023/dec/01/germany-scholz-government-handling-of-budget-crisis ↗
  12. Debortoli, S., Müller, O., Junglas, I. & Vom Brocke, J. (2016), ‘Text mining for information systems researchers: An annotated topic modeling tutorial’, Communications of the Association for Information Systems (CAIS) 39(1), 7. ↗
  13. Deutscher Bundestag (2025), ‘Deutscher Bundestag - Arbeit des Deutschen Bundestages’. URL: https://www.bundestag.de/services/faq/arbeit-244948 ↗
  14. Devika, M. D., Sunitha, C. & Ganesh, A. (2016), ‘Sentiment analysis: a comparative study on different approaches’, Procedia Computer Science 87, 44–49. find
  15. Dritsa, K., Thoma, A., Pavlopoulos, I. & Louridas, P. (2022), ‘A greek parliament proceedings dataset for computational linguistics and political analysis’, Advances in Neural Information Processing Systems 35, 28874–28888. ↗
  16. Eickhoff, M. & Neuss, N. (2017), Topic modelling methodology: Its use in information systems and other managerial disciplines, in ‘Proceedings of the 25th European Conference on Information Systems (ECIS)’, Research Papers, pp. 1327–1347. find
  17. Emmer, M. (2016), ‘Soziale medien in der politischen kommunikation’, Handbuch Soziale Medien pp. 81–99. ↗
  18. European Commission (2024), ‘Standard eurobarometer 101 – spring 2024: Public opinion in the european union’. URL: https://europa.eu/eurobarometer/surveys/detail/3216 ↗
  19. Fawcett, T. (2006), ‘An introduction to roc analysis’, Pattern recognition letters 27(8), 861–874. ↗
  20. Fiva, J. H., Nedregård, O. & Øien, H. (2025), ‘The norwegian parliamentary debates dataset’, Scientific Data 12(1), 4. find
  21. Güngör, O., Tiftikçi, M. & Sönmez, Ç. (2018), A corpus of grand national assembly of turkish parliament’s transcripts, in ‘Proceedings of the Eleventh International Conference on Language Resources and Evaluation’. find
  22. Hamann, J. (2016), "Das Volk" und "Die da oben": Misstrauen und Entfremdung in Europa, Vol. 17 of DGAP kompakt, Forschungsinstitut der Deutschen Gesellschaft für Auswärtige Politik e.V., Berlin. find
  23. Hofmann, S., Sommermann, C., Kraus, M., Zschech, P. & Rosenberger, J. (2024), Hate Speech and Sentiment of YouTube Video Comments From Public and Private Sources Covering the Israel-Palestine Conflict, in ‘Wirtschaftsinformatik 2024 Proceedings’. find
  24. Huang, J. & Ling, C. (2005), ‘Using AUC and accuracy in evaluating learning algorithms’, IEEE Transactions on Knowledge and Data Engineering 17(3), 299–310. ↗
  25. Jakubícek, M. & Kovár, V. (2010), Czechparl: Corpus of stenographic protocols from czech parliament., in ‘RASLAN’, pp. 41–46. find
  26. Kim, E. H.-J., Jeong, Y. K., Kim, Y., Kang, K. Y. & Song, M. (2016), ‘Topic-based content and sentiment analysis of ebola virus on twitter and in the news’, Journal of Information Science 42(6), 763–781. ↗
  27. Koltsova, O. & Koltcov, S. (2013), ‘Mapping the public agenda with topic modeling: The case of the russian livejournal’, Policy & Internet 5(2), 207–227. ↗
  28. Körber-Stiftung (2024), ‘Demokratie in der Krise Umfrage 2024’. URL: https://koerber-stiftung.de/projekte/demokratie-in-der-krise-umfrage-2024/ ↗
  29. Körber-Stiftung (2023), ‘Deutsche verlieren vertrauen in ihre demokratie’. URL: https://koerber-stiftung.de/presse/mitteilungen/deutsche-verlieren-vertrauen-in-ihre-demokratie/ ↗
  30. Laver, M., Benoit, K. & Garry, J. (2003), ‘Extracting policy positions from political texts using words as data’, American Political Science Review 97(2), 311–331. ↗
  31. Lilleberg, J., Zhu, Y. & Zhang, Y. (2015), Support vector machines and word2vec for text classification with semantic features, in ‘2015 IEEE 14th International Conference on Cognitive Informatics & Cognitive Computing (ICCI*CC)’, pp. 136–140. URL: https://ieeexplore.ieee.org/document/7259377/?arnumber=7259377 ↗
  32. Lorenz-Spreen, P., Oswald, L., Lewandowsky, S. & Hertwig, R. (2023), ‘A systematic review of worldwide causal and correlational evidence on digital media and democracy’, Nature human behaviour 7(1), 74–101. ↗
  33. Lu, B., Ott, M., Cardie, C. & Tsou, B. K. (2011), Multi-aspect sentiment analysis with topic models, in ‘2011 IEEE 11th international conference on data mining workshops’, IEEE, pp. 81–88. find
  34. Mansikkaniemi, A., Smit, P. & Kurimo, M. (2017), Automatic construction of the finnish parliament speech corpus, in ‘INTERSPEECH’, International Speech Communication Association (ISCA), pp. 3762–3766. find
  35. Marx, M., Schuth, A. et al. (2010), ‘Dutchparl. a corpus of parliamentary documents in dutch’, Proceedings Language Resources and Evaluation (LREC) pp pp. 3670–3677. find
  36. Meakem, A. (2024), ‘German government collapses amid uncertainties over u.s. election’. URL: https://foreignpolicy.com/2024/11/06/germany-coalition-government-collapse-scholz-lindner-budget-spending-trump/ ↗
  37. Nasution, A. H. & Onan, A. (2024), ‘Chatgpt label: Comparing the quality of human-generated and llm-generated annotations in low-resource language nlp tasks’, IEEE Access 12, 71876–71900. ↗
  38. Ogrodniczuk, M. (2018), Polish parliamentary corpus, in ‘Proceedings of the LREC 2018 workshop ParlaCLARIN: creating and using parliamentary corpora’, pp. 15–19. find
  39. Ogunleye, B., Maswera, T., Hirsch, L., Gaudoin, J. & Brunsdon, T. (2023), ‘Comparison of topic modelling approaches in the banking context’, Applied Sciences 13(2), 797. ↗
  40. Rauh, C. (2018), ‘Validating a sentiment dictionary for german political language—a workbench note’, Journal of Information Technology & Politics 15(4), 319–343. ↗
  41. Rauh, C. & Schwalbach, J. (2020), ‘The parlspeech v2 data set: Full-text corpora of 6.3 million parliamentary speeches in the key legislative chambers of nine representative democracies’, Harvard Dataverse 1(1). ↗
  42. Richter, F., Koch, P., Franke, O., Kraus, J., Warode, L., Kuruc, F., Heine, S. & Schöps, K. (2023), ‘Open discourse: towards the first fully comprehensive and annotated corpus of the parliamentary protocols of the german bundestag’, OSF Preprints . URL: https://doi.org/10.31235/osf.io/dx87u ↗
  43. Rosenberger, J., Wolfrum, L., Weinzierl, S., Kraus, M. & Zschech, P. (2025), ‘Career-BERT: Matching resumes to ESCO jobs in a shared embedding space for generic job recommendations’, Expert Systems with Applications 275, 127043. find
  44. Sahoo, B. P., Ashri, D. & Gulati, A. (2024), ‘Drivers of growth: Empirical analysis of ten largest economies’, The Indian Economic Journal p. 00194662241260815. ↗
  45. Shah, A. M., Yan, X., Qayyum, A., Naqvi, R. A. & Shah, S. J. (2021), ‘Mining topic and sentiment dynamics in physician rating websites during the early wave of the covid-19 pandemic: Machine learning approach’, International Journal of Medical Informatics 149, 104434. ↗
  46. Singh, A., Thakur, N. & Sharma, A. (2016), A review of supervised machine learning algorithms, in ‘2016 3rd international conference on computing for sustainable global development (INDIACom)’, Ieee, pp. 1310–1315. find
  47. Smelyakov, K., Karachevtsev, D., Kulemza, D., Samoilenko, Y., Patlan, O. & Chupryna, A. (2020), Effectiveness of preprocessing algorithms for natural language processing applications, in ‘2020 IEEE International Conference on Problems of Infocommunications. Science and Technology (PIC S&T)’, pp. 187–191. find
  48. Stine, R. A. (2019), ‘Sentiment analysis’, Annual review of statistics and its application 6(1), 287–308. find
  49. Storey, V. C. & Park, E. H. (2022), ‘An ontology of emotion process to support sentiment analysis’, Journal of the Association for Information Systems 23(4), 999–1036. ↗
  50. Tan, K. L., Lee, C. P. & Lim, K. M. (2023), ‘A survey of sentiment analysis: Approaches, datasets, and future research’, Applied Sciences 13(7), 4550. ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/patz-2025-analyzing-german-parliamentary-speeches-machine/