Elamur Research⌕
☰
home / papers / 2026 / One Joke to Rule them All? On the (Im)possibility of Generalizing Humor Detection

One Joke to Rule them All? On the (Im)possibility of Generalizing Humor Detection

Paper · 2026 · Mor Turgeman, Chen Shani and Dafna Shahaf

Summary

Computational humor research usually builds a separate model for each narrow genre, such as puns, memes or sarcasm. This paper tests whether that split is necessary: can a model that learns some kinds of humor recognize a kind it has never seen? The authors use four English binary humor-detection datasets: Amazon Questions, One Liners, Sarcasm Headlines and Reddit Dad Jokes. They fine-tune LLaMA-2-7B and Mistral-7B on one, two or three of these sources and then evaluate on every dataset, including fully held-out ones. Dad Jokes originally had no non-humorous examples, so the authors created matched negatives by having GPT-4 Turbo minimally rewrite jokes to strip out the punchline. The models do transfer to unseen humor types, but only partially and unevenly, and mixing training sources helps. The results suggest that humor genres share some underlying structure. That matters because new humor formats keep appearing online, and the authors release their data and code.

Key points

  • Data setup: Each dataset was cut down to 6,250 class-balanced examples and split 80/2/18 into train, validation and test. Every training set has 5,000 examples, including the multi-source ones. A 4-way domain classifier separated the datasets with 98–100% accuracy, which shows they are distinct.
  • Model choice: The two LLMs were picked because they score near chance (40–56%) on these tasks zero-shot. Any later gains can therefore be credited to fine-tuning and transfer.
  • Transfer results: On an unseen dataset, accuracy reaches up to 75% with Mistral, which generalizes better than LLaMA-2 overall. Training on more diverse sources adds 1.88–4.05%. In-domain accuracy barely drops, even when the target data makes up only a third of training.
  • Asymmetry: Dad Jokes is the strongest source for transfer to other datasets but the hardest target to transfer to. Headlines and One Liners are the reverse: easy to predict, but they help other datasets little.
  • Negative-example quality: A manual check of 3,000 generated negatives found that 2.63% failed to keep the original style or content. None of the checked outputs kept the punchline. Code is available at github.com/morturr/HumorTransferLearning.

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

🎓Stanford UniversityStanford · United States434 works · 314,419 citations🎓Hebrew University of JerusalemJerusalem · Israel65 works · 9,387 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 worksPsychology · 1,552 worksSocial Psychology · 306 works

Cite

Cite as Mor Turgeman, Chen Shani and Dafna Shahaf. 2026. One Joke to Rule them All? On the (Im)possibility of Generalizing Humor Detection. Proceedings of the 2nd Workshop on Computational Humor (CHum 2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.chum-1.1

@inproceedings{turgeman2026one,
  title = {{One Joke to Rule them All? On the (Im)possibility of Generalizing Humor Detection}},
  author = {Mor Turgeman, Chen Shani and Dafna Shahaf},
  booktitle = {Proceedings of the 2nd Workshop on Computational Humor (CHum 2026)},
  year = {2026},
  doi = {10.18653/v1/2026.chum-1.1},
  url = {https://doi.org/10.18653/v1/2026.chum-1.1}
}

References

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

  1. Romina Angeleri and Gabriella Airenti. 2014. The development of joke and irony understanding: a study with 3-to 6-year-old children. Canadian Journal of Experimental Psychology/Revue canadienne de psychologie expérimentale, 68(2):133. find
  2. Aseem Arora, Gaël Dias, Adam Jatowt, and Asif Ekbal. 2022. Transfer learning for humor detection by twin masked yellow muppets. In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Lan… find
  3. Salvatore Attardo. 2000. Irony as relevant inappropriateness. Journal of pragmatics, 32(6):793–826. find
  4. Salvatore Attardo. 2024. Linguistic theories of humor, volume 1. Walter de Gruyter GmbH & Co KG. find
  5. Sayak Autrin et al. 2023. Peft: Parameter-efficient fine-tuning. https://github.com/huggingface/peft. Accessed: 2025-07-20. ↗
  6. Alexander Baranov, Vladimir Kniazhevsky, and Pavel Braslavski. 2023. You told me that joke twice: A systematic investigation of transferability and robustness of humor detection models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13701–13715, Sing… find
  7. Hao Chen, Abdul Waheed, Xiang Li, Yidong Wang, Jindong Wang, Bhiksha Raj, and Marah I. Abdin. 2024. On the diversity of synthetic data and its impact on training large language models. Preprint, arXiv:2410.15226. \#\# Ethical considerations

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/turgeman-2026-one-joke-rule-them-all/