Elamur Research⌕
☰
home / papers / 2025 / German4All – A Dataset and Model for Readability-Controlled Paraphrasing in German

German4All – A Dataset and Model for Readability-Controlled Paraphrasing in German

Paper · 2025 · Miriam Anschütz, Thanh Mai Pham, Eslam Nasrallah et al.

Summary

German text simplification systems usually aim at a single fixed output level, but readers vary widely, from people with reading difficulties to experts. The authors address this with German4All, a corpus of Wikipedia paragraphs, each rewritten by GPT-4 into five aligned complexity levels. The levels are defined by target reader groups and aligned with OECD literacy proficiency levels rather than CEFR. The raw outputs passed through automatic format and language checks. A 150-sample test set was then hand-corrected by native speakers, and an LLM judge, validated against human ratings, filtered the large training split. With this data they fine-tune LoRA adapters on flan-t5-xl. The result is a compact, open model that rewrites a text to any requested level and runs on a 12GB consumer GPU. On the EASSE-DE benchmark it beats recent German simplification systems on nearly every metric. The work matters because it offers an open, privacy-friendly alternative to calling large API models, and it makes graded simplification and complexification possible in a language other than English.

Key points

  • Source texts come from the Cohere/wikipedia-22-12 dump, using popular German paragraphs of at most 80 words. Generation used gpt-4-turbo-2024-04-09 via the Batch API with a structured prompt and a single hand-written 1-shot example.
  • Of 26,665 generated samples, 26,337 survived format, OOV and language filtering. LLM-judge filtering then removed 814 more, leaving 25,459 in German4All-Main.
  • German4All-Corrected holds 150 manually fixed test samples with an extra Leichte Sprache paraphrase. German4All-Annotated stores original/GPT-4/corrected triplets labelled with six edit-operation types, such as removed_info and added_info.
  • Data quality was checked with feedback from 16 human participants. The LLM judge's criteria (e.g. information_addition, complexity_level) correlated significantly with the human annotations.
  • The distilled flan-t5-xl + LoRA model outperformed erlesen-leo-7b, mBART-DEplain-APA+web, mt5-simple-german-corpus and Simba on EASSE-DE in all criteria except compression. Its weak spots are occasional hallucinations at high levels and minor grammar slips.

Interesting ideas

  • I1. Each aligned paraphrase serves as a training input for every other target level, not only the original. This multiplies the data fivefold and exposes the model to many input styles.
  • I2. Cheap automatic detectors only flag suspicious samples, such as invalid JSON, runs of three or more spaCy OOV tokens, or foreign-language text detected by langdetect. A human then reviews just the flagged items, which keeps manual effort small while catching false positives.
  • I3. An LLM judge is trusted for filtering only after its scores are shown to correlate with human annotations on a hand-checked subset. Its raw annotations are also published so users can choose their own thresholds.
  • I4. Hand corrections are logged as typed edit operations, like "removed info" or "added info", paired with before and after versions. This turns quality control into reusable, analyzable data.
  • I5. The authors noticed that the judge rated absolute complexity instead of fitness for the requested level. Graded outputs therefore need relative, target-aware evaluation criteria.

For Mova

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

  • I1 (back): When training generation from the meaning graph, every surface variant of the same content could be treated as an interchangeable source and target. This would teach the realizer that style and content are separate knobs.
  • I2 (learn): Silver-data pipelines could flag only sentences where witness parsers disagree or tokens look anomalous, and send just those to an LLM-teacher or human tier.
  • I3 (learn): Any LLM teacher used to grade frames or trees should first be calibrated against a small gold set. Its verdicts should be stored as data rather than baked-in decisions, so filters can be re-tuned later.
  • I4 (learn): Logging repairs to parses or frames as typed operations would create a corpus from which self-induced repair rules can be mined and counted.
  • I5 (answer): When checking answers or explanations aimed at different school levels, evaluators should be asked whether the text suits the stated reader, not just how hard it is.

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 4The 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 works

Cite

Cite as Miriam Anschütz, Thanh Mai Pham, Eslam Nasrallah, Maximilian Müller, Cristian-George Craciun and Georg Groh. 2025. German4All – A Dataset and Model for Readability-Controlled Paraphrasing in German. Proceedings of the 18th International Natural Language Generation Conference.

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{anschutz2025german4all,
  title = {{German4All – A Dataset and Model for Readability-Controlled Paraphrasing in German}},
  author = {Miriam Anschütz, Thanh Mai Pham, Eslam Nasrallah, Maximilian Müller, Cristian-George Craciun and Georg Groh},
  booktitle = {Proceedings of the 18th International Natural Language Generation Conference},
  year = {2025},
  url = {https://aclanthology.org/2025.inlg-main.24/}
}

References

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

  1. Mina Almasi and Ross Kristensen-McLachlan. 2025. Alignment drift in CEFR-prompted LLMs for interactive Spanish tutoring. In Proceedings of the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA 2025), pages 70–88, Vienna, Austria. Association for Computational Linguistics. ↗
  2. Toni Amstad. 1978. Wie verständlich sind unsere Zeitungen? Ph.D. thesis, Universität Zürich. find
  3. Miriam Anschütz and Georg Groh. 2022. TUM social computing at GermEval 2022: Towards the significance of text statistics and neural embeddings in text complexity prediction. In Proceedings of the GermEval 2022 Workshop on Text Complexity Assessment of German Text, pages 21–26, Potsdam, Germany. Association for Computational Linguistics. find
  4. Miriam Anschütz, Joshua Oehms, Thomas Wimmer, Bartłomiej Jezierski, and Georg Groh. 2023. Language models for German text simplification: Overcoming parallel data scarcity through style-specific pre-training. In Findings of the Association for Computational Linguistics: ACL 2023, pages 1147–1158, Toronto, Canada. Association for Computational Linguistics. find
  5. Miriam Anschütz, Tringa Sylaj, and Georg Groh. 2024. Images speak volumes: User-centric assessment of image generation for accessible communication. In Proceedings of the Third Workshop on Text Simplification, Accessibility and Readability (TSAR 2024), pages 27–40, Miami, Florida, USA. Association for Computational Linguistics. find
  6. Hadi Asghari, Christopher Richter, Freya Hewett, Larissa Wunderlich, and Theresa Züger. 2024. Simba: Ai-powered text simplification. find
  7. Dennis Aumiller and Michael Gertz. 2022. Klexikon: A German dataset for joint summarization and simplification. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, pages 2693–2701, Marseille, France. European Language Resources Association. find
  8. Abdullah Barayan, Jose Camacho-Collados, and Fernando Alva-Manchego. 2025. Analysing zero-shot readability-controlled sentence simplification. In Proceedings of the 31st International Conference on Computational Linguistics, pages 6762–6781, Abu Dhabi, UAE. Association for Computational Linguistics. ↗
  9. Paul-Gerhard Barbu. 2024. Entwicklung einer anwendung zum Übersetzten von texten in leichter/einfacher sprache mithilfe von large language models (llms). Master’s thesis, Rosenheim Technical University of Applied Sciences. Advised and supervised by Prof. Dr. Gerd Beneken and Prof. Dr. Marcel Tilly. find
  10. Alison Chi, Li-Kuang Chen, Yi-Chen Chang, Shu-Hui Lee, and Jason S. Chang. 2023. Learning to paraphrase sentences to different complexity levels. Transactions of the Association for Computational Linguistics, 11:1332–1354. ↗
  11. Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Alex Castro-Ros, Marie Pellat, Kevin Robinson, Dasha Valter, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei. 2024. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1–53. find
  12. Council for Cultural Co-operation. Education Committee. Modern Languages Division Council of Europe. 2001. Common European framework of reference for languages: Learning, teaching, assessment. Cambridge University Press. find
  13. DIN-Normenausschuss Ergonomie. 2024. Einfache Sprache - Anwendung für das Deutsche - Teil 1: Sprachspezifische Festlegungen (plain language - application for the german language - part 1: Language-specific provisions, DIN 8581-1). find
  14. DIN-Normenausschuss Ergonomie. 2025. Empfehlungen für Deutsche Leichte Sprache (guidance for German Easy Language, DIN SPEC 33429). find
  15. Björn Engelmann, Christin Katharina Kreutz, Fabian Haak, and Philipp Schaer. 2024. ARTS: Assessing readability & text simplicity. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 14925–14942, Miami, Florida, USA. Association for Computational Linguistics. find
  16. Asma Farajidizaji, Vatsal Raina, and Mark Gales. 2024. Is it possible to modify text to a target readability level? an initial investigation using zero-shot large language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 9325–9339, Torino, Italia. ELRA and ICCL. find
  17. Leon Fruth, Robin Jegan, and Andreas Henrich. 2024. An approach towards unsupervised text simplification on paragraph-level for German texts. In Proceedings of the Workshop on DeTermIt! Evaluating Text Difficulty in a Multilingual Context @ LREC-COLING 2024, pages 77–89, Torino, Italia. ELRA and ICCL. find
  18. Yingqiang Gao, Kaede Johnson, David Froehlich, Luisa Carrer, and Sarah Ebling. 2025. Evaluating the effectiveness of direct preference optimization for personalizing german automatic text simplifications for persons with intellectual disabilities. Preprint, arXiv:2507.01479. ↗
  19. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, and Ava Spataru et al. 2024. The llama 3 herd of models. Preprint, arXiv:2407.21783. ↗
  20. Antje Heine. 2017. Deutsch als Fremd- und Zweitsprache – eine besondere Form Leichter Sprache? Überlegungen aus der Perspektive des Faches DaF/-DaZ”. „Leichte Sprache “im Spiegel theoretischer und angewandter Forschung, pages 401–414. find
  21. Freya Hewett, Hadi Asghari, and Manfred Stede. 2024. Elaborative simplification for German-language texts. In Proceedings of the 25th Annual Meeting of the Special Interest Group on Discourse and Dialogue, pages 29–39, Kyoto, Japan. Association for Computational Linguistics. find
  22. Matthew Honnibal, Ines Montani, Sofie Van Lan-deghem, and Adriane Boyd. 2020. spacy: Industrial-strength natural language processing in python. find
  23. Margot Madina, Itziar Gonzalez-Dios, and Melanie Siegel. 2023. Easy-to-read language resources and tools for three european languages. In Proceedings of the 16th International Conference on PErvasive Technologies Related to Assistive Environments, PETRA ’23, page 693–699, New York, NY, USA. Association for Computing Machinery. find
  24. Ali Malik, Stephen Mayhew, Christopher Piech, and Klinton Bicknell. 2024. From tarzan to Tolkien: Controlling the language proficiency level of LLMs for content generation. In Findings of the Association for Computational Linguistics: ACL 2024, pages 15670–15693, Bangkok, Thailand. Association for Computational Linguistics. find
  25. Jonathan Mallinson, Rico Sennrich, and Mirella Lapata. 2020. Zero-shot crosslingual sentence simplification. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5109–5126, Online. Association for Computational Linguistics. find
  26. Sabine Manning. 2024. Ki-tools für einfache sprache: Leistungen von 10 tools auf einen blick. Published 22.05.2024, last accessed 30.04.2025. find
  27. Babak Naderi, Salar Mohtaj, Kaspar Ensikat, and Sebastian Möller. 2019. Subjective assessment of text complexity: A dataset for german language. arXiv preprint. ↗
  28. OECD. 2013. Oecd skills outlook 2013: First results from the survey of adult skills. OECD Publishing, http://dx.doi.org/10.1787/ 9789264204256-en. ↗
  29. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations.
  30. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational Linguistics. ↗
  31. Lars Klöser, Mika Beele, Jan-Niklas Schagen, and Bodo Kraft. 2024. German text simplification: Finetuning large language models with semi-synthetic data. In Proceedings of the Fourth Workshop on Language Technology for Equality, Diversity, Inclusion, pages 63–72, St. Julian’s, Malta. Association for Computational Linguistics. find
  32. 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.
  33. Klaus Krippendorff. 2011. Computing krippendorff’s alpha-reliability. find
  34. Björn Plüster. 2023. Leolm: Igniting german-language llm research. find
  35. Joseph Liu, Yoonsoo Nam, Xinyue Cui, and Swabha Swayamdipta. 2025. Evaluation under imperfect benchmarks and ratings: A case study in text simplification. Preprint, arXiv:2504.09394. ↗
  36. Andreas Säuberli, Franz Holzknecht, Patrick Haller, Silvana Deilen, Laura Schiffl, Silvia Hansen-Schirra, and Sarah Ebling. 2024. Digital comprehensibility assessment of simplified texts among persons with intellectual disabilities. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, CHI ’24, New York, NY, USA. Association for Computing Machinery. find
  37. Christiane Maaß. 2020. Easy language–plain language– easy language plus: Balancing comprehensibility and acceptability. Frank & Timme, Berlin. find
  38. Thorben Schomacker, Michael Gille, Marina Tropmann-Frick, and Jörg von der Hülls. 2023. Data and approaches for German text simplification – towards an accessibility-enhanced communication. In Proceedings of the 19th Conference on Natural Language Processing (KONVENS 2023), pages 63–68, Ingolstadt, Germany. Association for Computational Lingustics. ↗
  39. Nicolas Spring, Annette Rios, and Sarah Ebling. 2021. Exploring German multi-level text simplification. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1339–1349, Held Online. IN-COMA Ltd. find
  40. Sanja Stajner. 2021. Automatic text simplification for social good: Progress and challenges. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2637–2652, Online. Association for Computational Linguistics. ↗
  41. Junyi Jessy Li. 2024. InfoLossQA: Characterizing and recovering information loss in text simplification. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, pages 4263–4294. ↗
  42. Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze Chen, and Chris Callison-Burch. 2016. Optimizing statistical machine translation for text simplification. Transactions of the Association for Computational Linguistics, 4:401–415. ↗
  43. Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. Bertscore: Evaluating text generation with bert. In International Conference on Learning Representations. ↗
  44. Regina Stodden. 2024a. EASSE-DE & EASSE-multi: Easier automatic sentence simplification evaluation for German & multiple languages. In Proceedings of the Third Workshop on Text Simplification, Accessibility and Readability (TSAR 2024), pages 107–116, Miami, Florida, USA. Association for Computational Linguistics. find
  45. Regina Stodden. 2024b. Reproduction & benchmarking of German text simplification systems. In Proceedings of the Workshop on DeTermIt! Evaluating Text Difficulty in a Multilingual Context @ LREC-COLING 2024, pages 1–15, Torino, Italia. ELRA and ICCL. find
  46. Regina Stodden, Omar Momen, and Laura Kallmeyer. 2023. DEplain: A German parallel corpus with intralingual translations into plain language for sentence and document simplification. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16441–16463, Toronto, Canada. Association for Computational Linguistics. find
  47. Thanh Mai Pham. 2024. German4all: Paraphrasing german texts to different complexity levels with gpt-generated synthetic data. Master’s thesis, Technical University of Munich. Advised and supervised by Miriam Anschütz and Georg Groh. find
  48. Vanessa Toborek, Moritz Busch, Malte Boßert, Christian Bauckhage, and Pascal Welke. 2023. A new aligned simple German corpus. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 11393–11412, Toronto, Canada. Association for Computational Linguistics. ↗
  49. Martina Toshevska and Sonja Gievska. 2025. Llm-based text style transfer: Have we taken a step forward? IEEE Access, 13:44707–44721. Jan Trienes, Sebastian Joseph, Jörg Schlötterer, Christin Seifert, Kyle Lo, Wei Xu, Byron C. Wallace, and Human opera- LLM Judge tion criterion Phi-4 Llama-3.3-70B-Instruct Qwen2.5-72B- Gemma-3-Instruct 27B-it corrected info -0.23 (0.007) -0.05 (0.566) -0.24 (0.006) -0.27 (0.001) 0.38 (0.000) 0.13 (0.126) 0.30 (0.001) 0.51 (0.000) 0.23 (0.008) 0.10 (0.231) 0.20 (0.022) 0.35 (0.000) 0.14 (0.101) 0.13 (0.130) 0.16 (0.059) 0.13 (0.153) -0.07 (0.401) 0.19 (0.031) 0.32 (0.000) 0.34 (0.000) content preservation removed info information addition hallucination information in origin addition added info information loss adjusted com- complexity plexity level *Table 4: Spearman correlation strength and p-values (in brackets) of human operations in the German4All-annotated dataset and the LLM judge criteria. Statistically significant correlations (p-values ≤ 0.01) are bolded.* ↗

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/anschutz-2025-german4all-dataset-model-readability-controlled/