Elamur Research⌕
☰
home / papers / 2026 / ICT-NLP at SemEval-2026 Task 1: Humor Generation via RAG-based Augmentation and Multi-LLM Internal-External Voting

ICT-NLP at SemEval-2026 Task 1: Humor Generation via RAG-based Augmentation and Multi-LLM Internal-External Voting

Paper · 2026 · Wutao Shen, Liyuan Huang, Jiawei He et al.

Summary

This system paper describes the ICT-NLP entry submitted to SemEval-2026 Task 1 (Humor Generation), competing in Subtask A's Chinese track. The subtask asks for jokes built either from a news headline or from a given keyword pair. The authors use one pipeline for both input types. First, Qwen-Max's web search on Alibaba Cloud Bailian serves as a lightweight RAG step that turns each headline into a news summary. Next, guided by Toplyn's comedy-writing theory, they pull one keyword from the headline and one from the summary, so the headline case becomes a keyword-pair problem. An LLM proposes three associations for each keyword. Several LLMs then pick a word combination and a humor style, such as pun or exaggeration, and each writes a candidate joke. The generating models vote anonymously on the candidates, and an external model settles cases without a majority. The submission tied for first place. The paper shows that collaborative LLM pipelines without fine-tuning can produce competitive humor. It also notes that humor is subjective and that LLM judges and human preferences often disagree.

Key points

  • Four generator and voter models were used: Qwen-Max, DeepSeek-V3.2, GLM-4.7 and Kimi-K2-Thinking. A joke is accepted directly if it wins more than 50% of the internal votes. Otherwise a third-party LLM arbitrates.
  • Each voter scored candidates on seven criteria: comprehensibility, offensiveness, joke detectability, funniness, creativity, relevance and fluency.
  • Out of 300 candidate groups, 71 winners received 4 votes, 120 received 3, and 109 went to arbitration. After arbitration, Kimi-K2-Thinking won 163 groups, DeepSeek-V3.2 won 98, GLM-4.7 won 39 and Qwen-Max won 0. Qwen-Max chose the Q&A Reversal style 127 times, which produced very short jokes.
  • In the official pairwise ELO evaluation, the system scored 1052 (95% CI [1009, 1094]). It was one of seven teams tied for 1st place.
  • In an internal check, 7 native Chinese judges rated 20 sampled groups. Only one group was judged not to be a valid joke, and the judges' preferences were spread across all four models.

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

🏛️Chinese Academy of SciencesBeijing · China194 works · 23,906 citations🔬Institute of Computing TechnologyBeijing · China51 works · 3,411 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

🧭 Topics 1Research subjects and languages the work is about.Language models & pretraining · 3,320 works

Cite

Cite as Wutao Shen, Liyuan Huang, Jiawei He, Lin Li and Jin Zhang. 2026. ICT-NLP at SemEval-2026 Task 1: Humor Generation via RAG-based Augmentation and Multi-LLM Internal-External Voting. Proceedings of the 20th International Workshop on Semantic Evaluation (2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.semeval-1.247

@inproceedings{shen2026ict,
  title = {{ICT-NLP at SemEval-2026 Task 1: Humor Generation via RAG-based Augmentation and Multi-LLM Internal-External Voting}},
  author = {Wutao Shen, Liyuan Huang, Jiawei He, Lin Li and Jin Zhang},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.247},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.247}
}

References

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

  1. Miriam Amin and Manuel Burghardt. 2020. A survey on approaches to computational humor generation. In Proceedings of the 4th Joint SIGHUM Workshop on Computational Linguistics for Cultural Heritage, Social Sciences, Humanities and Literature, pages 29–41. find
  2. Santiago Castro, Luis Chiruzzo, Santiago Góngora, Salar Rahili, Naihao Deng, Ignacio Sastre, Victoria Amoroso, Guillermo Rey, Aiala Rosá, Guillermo Moncecchi, J. A. Meaney, Juan José Prada, and Rada Mihalcea. 2026. SemEval-2026 Task 1: MWA-HAHA, Models Write Automatic Humor And Humans Annotate. In P… find
  3. Yuetian Chen, Bowen Shi, and Mei Si. 2023. Prompt to gpt-3: step-by-step thinking instructions for humor generation. arXiv preprint arXiv:2306.13195. — [arXiv:2306.13195](https://arxiv.org/abs/2306.13195) find
  4. DeepSeek-AI. 2025. Deepseek-v3.2: Pushing the frontier of open large language models.
  5. Shivam Dubey. 2025. Humorplansearch: Structured planning and hucot for contextual ai humor. arXiv preprint arXiv:2508.11429. — [arXiv:2508.11429](https://arxiv.org/abs/2508.11429) find
  6. Evgenii Evstafev. 2025. Optimizing humor generation in large language models: Temperature configurations and architectural trade-offs. arXiv preprint arXiv:2504.02858. — [arXiv:2504.02858](https://arxiv.org/abs/2504.02858) find
  7. Ruiqi He, Yushu He, Longju Bai, Jiarui Liu, Zhenjie Sun, Zenghao Tang, He Wang, Hanchen Xia, Rada Mihalcea, and Naihao Deng. 2025. Chumor 2.0: Towards better benchmarking chinese humor understanding from (ruo zhi ba). In Findings of the Association for Computational Linguistics: ACL 2025, pages 2179… find
  8. EunJeong Hwang, Peter West, and Vered Shwartz. 2025. Bottlehumor: Self-informed humor explanation using the information bottleneck principle. In Findings of the Association for Computational Linguistics: ACL 2025, pages 22611–22632. find
  9. Sophie Jentzsch and Kristian Kersting. 2023. Chatgpt is fun, but it is not funny! humor is still challenging large language models. In Proceedings of the 13th Workshop on Computational Approaches to Subjectivity, Sentiment, & Social Media Analysis, pages 325–340. find
  10. Antonios Kalloniatis and Panagiotis Adamidis. 2024. Computational humor recognition: a systematic literature review. Artificial Intelligence Review, 58(2):43. find
  11. Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, and 1 others. 2025. Openai gpt-5 system card. arXiv preprint arXiv:2601.03267.
  12. GLM Team, Aohan Zeng, Xin Lv, Qinkai Zheng, Zhenyu Hou, Bin Chen, Chengxing Xie, Cunxiang Wang, Da Yin, Hao Zeng, Jiajie Zhang, Kedong Wang, Lucen Zhong, Mingdao Liu, Rui Lu, Shulin Cao, Xiaohan Zhang, Xuancheng Huang, Yao Wei, and 152 others. 2025a. Glm-4.5: Agentic, reasoning, and coding (arc) fou…
  13. Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025b. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534.
  14. Alexey Tikhonov and Pavel Shtykovskiy. 2024. Humor mechanics: Advancing humor generation with multi-step reasoning. arXiv preprint arXiv:2405.07280. — [arXiv:2405.07280](https://arxiv.org/abs/2405.07280) find
  15. Joe Toplyn. 2014. Comedy writing for late-night tv: How to write monologue jokes. Desk Pieces, Sketches, Parodies, Audience Pieces, Remotes, and Other Short-Form Comedy. find
  16. Joe Toplyn. 2023. Witscript 3: A hybrid ai system for improvising jokes in a conversation. Preprint, arXiv:2301.02695. — [arXiv:2301.02695](https://arxiv.org/abs/2301.02695) find
  17. Mor Turgeman, Chen Shani, and Dafna Shahaf. 2025. One joke to rule them all? on the (im) possibility of generalizing humor. arXiv preprint arXiv:2508.19402. — [arXiv:2508.19402](https://arxiv.org/abs/2508.19402) find
  18. Han Wang, Yilin Zhao, Dian Li, Xiaohan Wang, sin-badliu, Xuguang Lan, and Hui Wang. 2025. Innovative thinking, infinite humor: Humor research of large language models through structured thought leaps. In The Thirteenth International Conference on Learning Representations. find
  19. Orion Weller and Kevin Seppi. 2019. Humor detection: A transformer gets the last laugh. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3621–3625.
  20. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388.
  21. Zhenghan Yu, Xinyu Hu, and Xiaojun Wan. 2025. Cfun-model: A" funny" language model capable of chinese humor generation and processing. arXiv preprint arXiv:2503.20417. — [arXiv:2503.20417](https://arxiv.org/abs/2503.20417) find
  22. Zhiwei Yu, Jiwei Tan, and Xiaojun Wan. 2018. A neural approach to pun generation. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1650–1660, Melbourne, Australia. Association for Computational Linguistics. find
  23. Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. 2024. Let’s think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti… find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/shen-2026-ict-nlp-semeval-2026-task/