Elamur Research⌕
☰
home / papers / 2025 / CourtPressGER: A German Court Decision to Press Release Summarization Dataset

CourtPressGER: A German Court Decision to Press Release Summarization Dataset

Paper · 2025 · Sebastian Nagl, Mohamed Elganayni, Melanie Pospisil et al.

Summary

German federal courts publish press releases that explain rulings to journalists and ordinary citizens. Earlier legal NLP work on German focused on technical headnotes for lawyers. This paper presents CourtPressGER, a corpus of about 6.4k triples. Each triple holds a ruling, the press release a human wrote for it, and an LLM-written prompt that describes the target release's length, audience and topics.

The authors test six open and commercial LLMs of different sizes. Small-context models get rulings that are too long through hierarchical chunk-and-merge summarization. They also fine-tune Teuken-7B in two stages for this task. Outputs are scored with overlap metrics, embedding metrics, factual-consistency checks, an LLM judge and expert rankings.

Large models such as GPT-4o and Llama-3-70B write strong drafts and lose little when run hierarchically. Small models need the hierarchical setup to cope. The fine-tuned 7B model still trails far behind the large ones. Experts always ranked the human-written releases first.

The resource opens benchmarking of citizen-oriented legal summarization in German. It also shows that LLM-judge scores agree closely with expert rankings.

Key points

  • The corpus has 6,432 decision–press release pairs from six German federal courts: BAG, BFH, BGH, BSG, BVerfG and BVerwG. They cover 1995–2023 and are split by date into 4,643 train, 744 validation and 1,045 test pairs.
  • Measured with the EuroBERT tokenizer, a ruling averages about 10.8k tokens (std ≈10.7k) and a release about 1.4k tokens (std ≈955).
  • Per-pair prompts were produced with Claude Sonnet 3.7. The models tested were GPT-4o, Llama-3-70B, Llama-3-8B, Teuken-7B, EuroLLM-9B and Mistral-7B.
  • Teuken-7B was tuned in two stages using stacked SFT:
    • first on roughly 35k chunk–instruction pairs, with reference summaries written by Llama-3-70B-Instruct;
    • then on mapping the top-level summary plus the prompt to the gold release.
  • LLM-judge scores closely matched human rankings: Spearman ρ = −0.939, Kendall τ = −0.822, Pearson r = −0.987. This rests on only 10 configurations and 14 human-rated cases. QAGS and FactCC were judged unreliable for German legal text.

Interesting ideas

  • I1. Splitting the data by date rather than at random shows the effect of real drift, such as changing press-office writers, and keeps test scores from looking too good.
  • I2. Each example gets its own generated prompt describing the length, audience and coverage of the reference, so outputs are compared against matching targets rather than one generic instruction.
  • I3. Chunk summaries are merged level by level in a tree instead of updating one summary chunk after chunk, which the authors note keeps long outputs more coherent.
  • I4. The LLM judge is trusted only after its scores are checked against expert rankings with several rank correlations.
  • I5. Consistency metrics falsely penalize a reference when it contains true context absent from the source, and English-built metrics may misread German.

For Mova

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

  • I1 (learn): Time- or source-based held-out splits for silver treebanks and QA sets would reveal whether repair rules and parsers survive shifts in style.
  • I2 (back): Generation from the meaning graph could be tested against a short per-text spec (length, register, which entities to mention), so that text → graph → text round trips are judged fairly.
  • I3 (world): A long text's world model could be built per section and then merged in a tree, reconciling entities and event order at each merge.
  • I4 (learn): Each LLM teacher tier could be calibrated this way against a small hand-checked gold set before its labels count as silver data.
  • I5 (answer): QA checks on tales and tests should separate facts that can be grounded in the text from licit outside knowledge, so correct answers drawn from world knowledge are not marked as hallucinations.

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

🎓Technical University of MunichMunich · Germany55 works · 48,474 citations

 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 worksPolitical Science and International Relations · 182 works

🧭 Topics 2Research subjects and languages the work is about.Summarization · 232 worksLegal AI · 134 works

Cite

Cite as Sebastian Nagl, Mohamed Elganayni, Melanie Pospisil and Matthias Grabmair. 2025. CourtPressGER: A German Court Decision to Press Release Summarization Dataset. arXiv:2512.09434.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2512.09434 · arXiv 2512.09434

@misc{nagl2025courtpressger,
  title = {{CourtPressGER: A German Court Decision to Press Release Summarization Dataset}},
  author = {Sebastian Nagl, Mohamed Elganayni, Melanie Pospisil and Matthias Grabmair},
  year = {2025},
  doi = {10.48550/arxiv.2512.09434},
  eprint = {2512.09434},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2512.09434}
}

References

31 references, 7 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. C. Grover, B. Hachey, I. Hughson, B. Place, The holj corpus: Supporting summarisation of legal texts, in: In Proceedings of the 5th International Workshop on Linguistically Interpreted Corpora, 2004. find
  2. S. Polsley, P. Jhunjhunwala, R. Huang, Casesummarizer: a system for automated summarization of legal texts, in: Proceedings of COLING 2016, the 26th international conference on Computational Linguistics: System Demonstrations, 2016, pp. 258–262. find
  3. I. Chalkidis, D. Kampas, Deep learning in law: early adaptation and legal word embeddings trained on large corpora, Artificial Intelligence and Law 27 (2019) 171–198. find
  4. T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, Y. Artzi, BERTScore: Evaluating Text Generation with BERT, 2020. doi: 10.48550/arXiv.1904.09675 . arXiv:1904.09675 . ↗
  5. D. Aumiller, A. Chouhan, M. Gertz, Eur-lex-sum: A multi-and cross-lingual dataset for long-form summarization in the legal domain, arXiv preprint arXiv:2210.13448 (2022). ↗
  6. A. Kanapala, S. Pal, R. Pamula, Text summarization from legal documents: a survey, Artificial Intelligence Review 51 (2019) 371–402. find
  7. M. Akter, E. Cano, E. Weber, D. Dobler, I. Habernal, A comprehensive survey on legal summarization: Challenges and future directions, arXiv preprint arXiv:2501.17830 (2025). ↗
  8. Y. Chang, K. Lo, T. Goyal, M. Iyyer, BooookScore: A systematic exploration of book-length summarization in the era of LLMs, 2024. doi: 10.48550/arXiv.2310.00785 . arXiv:2310.00785 . ↗
  9. A. Deroy, K. Ghosh, S. Ghosh, Ensemble methods for improving extractive summarization of legal case judgements, Artificial Intelligence and Law 32 (2024) 231–289. find
  10. B. Steffes, P. Rataj, L. Burger, L. Roth, On evaluating legal summaries with rouge, in: Proceedings of the Nineteenth International Conference on Artificial Intelligence and Law, 2023, pp. 457–461. find
  11. A. Kornilova, V. Eidelman, Billsum: A corpus for automatic summarization of us legislation, arXiv preprint arXiv:1910.00523 (2019). ↗
  12. Z. Shen, K. Lo, L. Yu, N. Dahlberg, M. Schlanger, D. Downey, Multi-lexsum: Real-world summaries of civil rights lawsuits at multiple granularities, Advances in Neural Information Processing Systems 35 (2022) 13158–13173. find
  13. D. d. V. Feijo, V. P. Moreira, Improving abstractive summarization of legal rulings through textual entailment, Artificial intelligence and law 31 (2023) 91–113. find
  14. I. Glaser, S. Moser, F. Matthes, Summarization of German Court Rulings, in: Proceedings of the Natural Legal Language Processing Workshop 2021, Association for Computational Linguistics, Punta Cana, Dominican Republic, 2021, pp. 180–189. doi: 10.18653/v1/2021.nllp-1.19 . ↗
  15. L. Rolshoven, V. Rasiah, S. B. Bose, M. Stürmer, J. Niklaus, Unlocking Legal Knowledge: A Multilingual Dataset for Judicial Summarization in Switzerland, 2024. doi: 10.48550/arXiv.2410. 13456 . arXiv:2410.13456 . ↗
  16. L. Manor, J. J. Li, Plain english summarization of contracts, arXiv preprint arXiv:1906.00424 (2019). ↗
  17. O. Salaün, A. Troussel, S. Longhais, H. Westermann, P. Langlais, K. Benyekhlef, Conditional abstractive summarization of court decisions for laymen and insights from human evaluation, in: Legal Knowledge and Information Systems, IOS Press, 2022, pp. 123–132. find
  18. M. Elaraby, D. Litman, Arglegalsumm: Improving abstractive summarization of legal documents with argument mining, 2022. URL: https://arxiv.org/abs/2209.01650. arXiv:2209.01650 . ↗
  19. A. Shukla, P. Bhattacharya, S. Poddar, R. Mukherjee, K. Ghosh, P. Goyal, S. Ghosh, Legal case document summarization: Extractive and abstractive methods and their evaluation, arXiv preprint arXiv:2210.07544 (2022). ↗
  20. G. Moro, L. Ragazzi, Semantic self-segmentation for abstractive summarization of long documents in low-resource regimes, in: Proceedings of the AAAI conference on artificial intelligence, volume 36, 2022, pp. 11085–11093. find
  21. T. Y. S. S. Santosh, Y. Farag, M. Grabmair, Coperlex: Content planning with event-based representations for legal case summarization, 2025. URL: https://arxiv.org/abs/2501.14112. arXiv:2501.14112 . ↗
  22. T. Santosh, C. Weiss, M. Grabmair, Lexsumm and lext5: Benchmarking and modeling legal summarization tasks in english, arXiv preprint arXiv:2410.09527 (2024). ↗
  23. T. Santosh, V. Venkatkrishna, S. Ghosh, M. Grabmair, Beyond borders: Investigating cross-jurisdiction transfer in legal case summarization, arXiv preprint arXiv:2403.19317 (2024). ↗
  24. A. Wang, K. Cho, M. Lewis, Asking and Answering Questions to Evaluate the Factual Consistency of Summaries, 2020. doi: 10.48550/arXiv.2004.04228 . arXiv:2004.04228 . ↗
  25. W. Kryściński, B. McCann, C. Xiong, R. Socher, Evaluating the Factual Consistency of Abstractive Text Summarization, 2019. doi: 10.48550/arXiv.1910.12840 . arXiv:1910.12840 . ↗
  26. H. Xu, J. Savelka, K. D. Ashley, Toward summarizing case decisions via extracting argument issues, reasons, and conclusions, in: Proceedings of the eighteenth international conference on artificial intelligence and law, 2021, pp. 250–254. find
  27. A. Deroy, K. Ghosh, S. Ghosh, Applicability of Large Language Models and Generative Models for Legal Case Judgement Summarization, 2024. doi: 10.48550/arXiv.2407.12848 . arXiv:2407.12848 . ↗
  28. A. Pareja, N. S. Nayak, H. Wang, K. Killamsetty, S. Sudalairaj, W. Zhao, S. Han, A. Bhandwaldar, G. Xu, K. Xu, L. Han, L. Inglis, A. Srivastava, Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs, 2024. doi: 10.48550/arXiv.2412.13337 . arXiv:2412.13337 . ↗
  29. C.-Y. Lin, ROUGE: A Package for Automatic Evaluation of Summaries, ACL - Text Summarization Branches Out (2004).
  30. K. Papineni, S. Roukos, T. Ward, W.-J. Zhu, BLEU: A method for automatic evaluation of machine translation, in: Proceedings of the 40th Annual Meeting on Association for Computational Linguistics - ACL ’02, Association for Computational Linguistics, Philadelphia, Pennsylvania, 2002, p. 311. doi: 10.3115/1073083.1073135 . ↗
  31. S. Banerjee, A. Lavie, METEOR: An Automatic Metric for MT Evaluation with Improved Correlation with Human Judgments, Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization (2005).

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/nagl-2025-courtpressger-german-court-decision-press/