Context-Aware Content Moderation for German Newspaper Comments

Paper · 2025 · Felix Krejca, Tobias Kietreiber, Alexander Buchelt et al.

Summary

Newspaper comment sections need moderation at scale. Human moderators are exposed to harmful content, and regulation such as the EU DSA increases the pressure on platforms. Most work on German moderation looks only at the comment text and ignores the platform metadata around it. The authors frame moderation as a binary choice: keep the post online or take it offline. They use the One Million Posts Corpus from Der Standard and test whether context improves the decision. The context signals are the article title, the article's topic path, and the share of a user's earlier posts that stayed online. They compare base and advanced LSTM and CNN models, classic baselines (logistic regression, naive Bayes) and zero-shot GPT-3.5 Turbo prompts with and without context. Context clearly helps the self-trained networks, and the user-history signal helps most. The GPT prompts gain nothing from context and score lowest. The work suggests that small context-aware models can be a practical choice for platform-specific moderation.

Key points

Data: One Million Posts Corpus, with 1,000,000 posts labeled online/offline by Der Standard moderators and 11,773 posts with richer multi-labels.
Best model: adv_LSTM_Title_path_ratio, with accuracy 0.733 and AUROC 0.809. Logistic regression reached 0.663 accuracy and naive Bayes 0.678.
GPT-3.5 Turbo zero-shot scored 0.606–0.635 accuracy and often gave a "missing answer" (wrong format or refusal). It showed high recall but low precision.
The Online Ratio user-history feature gave the largest gain, while the topic Path added little. Advanced LSTM and CNN models performed about the same.
Compared with earlier work, Assenmacher et al.'s BERT has a higher AUROC (0.914). Against Yadav & Milde, the best model has better F1, precision and recall but lower accuracy.

Interesting ideas

I1. The authors add context in stages (text only, then plus title, then plus path, then plus user ratio), which shows exactly how much each signal contributes.
I2. Failed or refused LLM outputs are counted and reported as a separate "missing answer" figure instead of being silently dropped.
I3. Adding context to a prompt raised cost but not accuracy, so more input is not automatically more understanding.
I4. A user's history (the share of earlier posts kept online) served as a strong prior next to the text itself.
I5. A model with high recall and low precision was seen as better suited to supporting humans than to deciding alone.

For Mova

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

I1 (learn): Mova's evaluations could use the same incremental design, turning on discourse context, coreference or knowledge one layer at a time to measure each layer's real value.
I2 (learn): LLM teacher tiers should log unparseable or refused annotations as their own error class, so the reliability of silver data is visible.
I3 (frames): pragmatic judgments such as irony or speech acts could pass only the context that structured analysis shows to be relevant, rather than whole surrounding documents.
I4 (world): the text world could store per-speaker or per-character priors (past behaviour, reliability) as explicit entity states that feed into later inferences.
I5 (answer): readers could report confidence profiles and send recall-heavy, uncertain answers to review instead of committing to them.

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: University of Applied Sciences St Pölten.

Fields, topics and concepts

📚 Shelves 4Computer Science · 16,667 works Artificial Intelligence · 16,194 works Social Sciences · 1,772 works Sociology and Political Science · 577 works

Cite

Cite as Felix Krejca, Tobias Kietreiber, Alexander Buchelt and Sebastian Neumaier. 2025. Context-Aware Content Moderation for German Newspaper Comments. arXiv:2505.20963.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2505.20963 · arXiv 2505.20963

@misc{krejca2025context,
  title = {{Context-Aware Content Moderation for German Newspaper Comments}},
  author = {Felix Krejca, Tobias Kietreiber, Alexander Buchelt and Sebastian Neumaier},
  year = {2025},
  doi = {10.48550/arxiv.2505.20963},
  eprint = {2505.20963},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2505.20963}
}

References

7 references, 3 of them with a card of ours. ↗ leads to the work.

1. Alkomah, F., Ma, X.: A literature review of textual hate speech detection methods and datasets. Information 13(6) (2022). https://doi.org/10.3390/info13060273 2. Assenmacher, D., Niemann, M., Müller, K., Seiler, M.V., Riehle, D.M., Trautmann, H.: RP-Mod & RP-Crowd: Moderator- and Crowd-Annotated German News Comment Datasets (2021). https://doi.org/10.5281/zenodo.5242915, https://doi.org/10.5281/zenodo.5242915, [dataset] 3. Chiu, K.L., Collins, A., Alexander, R.: Detecting hate speech with gpt-3 (2022), https://arxiv.org/abs/2103.12407 4. Dehghan, S., Yanikoglu, B.: Evaluating ChatGPT‘s ability to detect hate speech in Turkish tweets. In: Proceedings of the 7th Workshop on Challenges and Applications of Automated Extraction of Socio-political Events from Text (2024). pp. 54–59. Association for Computational Linguistics, St. ↗
Julians, Malta (2024), https://aclanthology.org/2024.case-1.6/ ↗
Context-Aware Content Moderation for German Newspaper Comments 5. Fortuna, P., Nunes, S.: A survey on automatic detection of hate speech in text. ACM Comput. Surv. 51(4) (2018). https://doi.org/10.1145/3232676 6. Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735–1780 (1997). https://doi.org/10.1162/neco.1997.9.8.1735 7. Horta Ribeiro, M., Cheng, J., West, R.: Automated content moderation increases adherence to community guidelines. In: Proceedings of the ACM Web Conference 2023. p. 2666–2676. WWW ’23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3543507.3583275 8. Jaremko, J., Gromann, D., Wiegand, M.: Revisiting implicitly abusive language detection: Evaluating llms in zero-shot and few-shot settings. In: Proceedings of the 31st International Conference on Computational Linguistics. pp. 3879–3898 (2025) 9. Jose, M., Anthony, J., Joseph, J.V., Thomas, J., Thomas, S.B.: A review of machine learning and deep learning approaches for offensive text detection. International Journal on Emerging Research Areas (IJERA) 04(02), 47–50 (2025). https://doi. org/10.5281/zenodo.14651005, https://doi.org/10.5281/zenodo.14651005 10. Joulin, A., Grave, E., Bojanowski, P., Mikolov, T.: Bag of tricks for efficient text classification. In: Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers. pp. 427–431. Association for Computational Linguistics, Valencia, Spain (2017), https://aclanthology.org/E17-2068/ 11. Kang, Y., Cai, Z., Tan, C.W., Huang, Q., Liu, H.: Natural language processing (nlp) in management research: A literature review. Journal of Management Analytics 7(2), 139–172 (2020). https://doi.org/10.1080/23270012.2020.1756939 12. Keller, M.E., Auch, M., Döschl, A., Vlk, F., Quernheim, J., Hartmann, M., Mandl, P., Kaul, A., Franz, M., Greguš, M., Kotsis, G., Delir Haghighi, P., Khalil, I.: Hocon34k: A corpus of hate speech in online comments from german newspapers. In: Information Integration and Web Intelligence, pp. 212–226. ↗
Springer Nature Switzerland (2025). https://doi.org/10.1007/978-3-031-78090-5_18, https://doi. org/10.1007/978-3-031-78090-5_18 13. Krenn, B., Petrak, J., Kubina, M., Burger, C.: Germs-at: A sexism/misogyny dataset of forum comments from an austrian online newspaper. In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024). pp. 7728–7739 (2024) 14. Li, L., Fan, L., Atreja, S., Hemphill, L.: “hot” chatgpt: The promise of chatgpt in detecting and discriminating hateful, offensive, and toxic comments on social media. ACM Trans. Web 18(2) (2024). https://doi.org/10.1145/3643829 15. Lu, J., Ma, K., Wang, K., Xiao, K., Lee, R.K.W., Xu, B., Yang, L., Lin, H.: Unveiling the capabilities of large language models in detecting offensive language with annotation disagreement. https://doi.org/10.48550/arxiv.2502.06207 (2025), arXiv:2502.06207 16. Malik, J.S., Qiao, H., Pang, G., van den Hengel, A.: Deep learning for hate speech detection: a comparative study. International Journal of Data Science and Analytics pp. 1–16 (2024) 17. Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word representations in vector space (2013), https://arxiv.org/abs/1301.3781 18. Mnassri, K., Farahbakhsh, R., Chalehchaleh, R., Rajapaksha, P., Jafari, A.R., Li, G., Crespi, N.: A survey on multi-lingual offensive language detection. PeerJ. Computer Science 10, e1934–e1934 (2024). https://doi.org/10.7717/peerj-cs.1934, https://doi.org/10.7717/peerj-cs.1934 ↗
F. Krejca, et al. 19. Mullah, N.S., Zainon, W.M.N.W.: Advances in machine learning algorithms for hate speech detection in social media: A review. IEEE Access 9, 88364–88376 (2021). https://doi.org/10.1109/ACCESS.2021.3089515 20. OpenAI: Openai api (2025), https://mistral.ai/en/news/announcing-mistral-7b 21. Pachinger, P., Goldzycher, J., Planitzer, A.M., Kusa, W., Hanbury, A., Neidhardt, J.: Austrotox: A dataset for target-based austrian german offensive language detection. arXiv preprint arXiv:2406.08080 (2024), https://arxiv.org/abs/2406.08080 22. Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word representation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 1532–1543 (2014) 23. Raschka, S., Liu, Y.H., Mirjalili, V.: Machine Learning with PyTorch and Scikit-Learn: Develop machine learning and deep learning models with Python. ↗
Packt Publishing Ltd (2022) 24. Saumya, S., Kumar, A., Singh, J.P.: Filtering offensive language from multilingual social media contents: A deep learning approach. Engineering Applications of Artificial Intelligence 133(Part B), 108159 (2024). https://doi.org/10.1016/j.engappai. 2024.108159, https://doi.org/10.1016/j.engappai.2024.108159 25. Schabus, D., Skowron, M.: Academic-industrial perspective on the development and deployment of a moderation system for a newspaper website. In: Proceedings of the eleventh international conference on language resources and evaluation (2018) 26. Schabus, D., Skowron, M., Trapp, M.: One million posts: A data set of german online discussions. In: Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval. p. 1241–1244. SIGIR ’17, New York, NY, USA (2017). https://doi.org/10.1145/3077136.3080711 27. Spence, R., Bifulco, A., Bradbury, P., Martellozzo, E., DeMarco, J.: The psychological impacts of content moderation on content moderators: A qualitative study. Cyberpsychology: Journal of Psychosocial Research on Cyberspace 17(4) (2023) 28. Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I.: Attention is all you need. arXiv preprint arXiv:1706.03762 (2017), https://doi.org/10.48550/arxiv.1706.03762 29. Vidgen, B., Derczynski, L.: Directions in abusive language training data, a systematic review: Garbage in, garbage out. PLOS ONE 15(12), 1–32 (2021). https://doi.org/10.1371/journal.pone.0243300 30. Waseem, Z., Davidson, T., Warmsley, D., Weber, I.: Understanding abuse: A typology of abusive language detection subtasks. In: Proceedings of the First Workshop on Abusive Language Online. pp. 78–84. Association for Computational Linguistics, Vancouver, BC, Canada (2017). https://doi.org/10.18653/v1/W17-3012 31. Yadav, A., Milde, B.: forumBERT: Topic adaptation and classification of contextualized forum comments in German. In: Proceedings of the 17th Conference on Natural Language Processing (KONVENS 2021). pp. 193–202. ↗
Düsseldorf, Germany (2021), https://aclanthology.org/2021.konvens-1.17/ 32. Zampieri, M., Malmasi, S., Nakov, P., Rosenthal, S., Farra, N., Kumar, R.: SemEval-2019 task 6: Identifying and categorizing offensive language in social media (OffensEval). In: Proceedings of the 13th International Workshop on Semantic Evaluation. pp. 75–86. Association for Computational Linguistics, Minneapolis, Minnesota, USA (2019). https://doi.org/10.18653/v1/S19-2010 ↗

Updated 30 October 2026

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/krejca-2025-context-aware-content-moderation-german/