---
title: "SLPGFJWUInsa at SemEval-2026 Task 1: Enhancing Linguistic"
description: "This SemEval-2026 Task 1 (MWAHAHA) system addresses Subtask A: writing English jokes under constraints. 13 references."
authors: ["Insa Abbas", "Sadaf Abdul Rauf"]
year: 2026
venue: "Proceedings of the 20th International Workshop on Semantic Evaluation (2026)"
source: "https://aclanthology.org/2026.semeval-1.231/"
doi: "10.18653/v1/2026.semeval-1.231"
licence: "CC BY 4.0 (ACL Anthology)"
licence_url: "https://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/abbas-2026-slpgfjwuinsa-semeval-2026-task-1/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/abbas-2026-slpgfjwuinsa-semeval-2026-task-1/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/43/) / SLPGFJWUInsa at SemEval-2026 Task 1: Enhancing Linguistic Creativity for English Text-Based Humor

# SLPGFJWUInsa at SemEval-2026 Task 1: Enhancing Linguistic Creativity for English Text-Based Humor

**Authors:** Insa Abbas and Sadaf Abdul Rauf · **Year:** 2026 · **Venue:** [Proceedings of the 20th International Workshop on Semantic Evaluation (2026)](https://md.research.elamur.ai/free/en/venues/semeval/) ([SemEval 2026](https://md.research.elamur.ai/free/en/venues/semeval/2026/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY 4.0 (ACL Anthology)](https://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.semeval-1.231/) · [PDF](https://aclanthology.org/2026.semeval-1.231.pdf) · [DOI 10.18653/v1/2026.semeval-1.231](https://doi.org/10.18653/v1/2026.semeval-1.231) · [OpenAlex](https://openalex.org/W7166859713)

**Citations:** 0 citations (OpenAlex, 2026-10-14)

## Summary
This SemEval-2026 Task 1 (MWAHAHA) system addresses Subtask A: writing English jokes under constraints. Each input is either a rare word pair that the joke must use or a news headline to satirize, and every output must begin with the marker JOKE:. The authors built a dataset of 12,118 input–joke pairs in four stages. ChatGPT-4o and Gemini 1.5 Pro drafted the candidates, three annotators screened them, the text was cleaned, and a final pass checked the format. They then adapted the small Phi-2 (2.7B) model with QLoRA and enlarged the training side with WordNet synonym swaps and three rule-based input rewrites. Checkpoints saved at 2, 4, 6, 8 and 10 epochs were compared. Epoch 4 was submitted because dev perplexity and training loss begin to diverge after that point. The system placed first in the human evaluation with a score of 1080. The result suggests that a compact model can be competitive at creative generation if it gets curated data, careful checkpoint choice and strict output templates.

## Key points
- Data pipeline: about 8,000 prompts went to Gemini and about 4,000 to ChatGPT-4o. Gemini was used more heavily because its jokes stuck to the constraints more often. Manual filtering (Cohen's κ = 0.74) discarded roughly 34% of candidates. The split is 90/10 train/dev, with 1,847 unique anchors.
- Model setup:
  - Phi-2 with 4-bit NF4 QLoRA.
  - LoRA rank 16 and α = 32, applied only to Wqkv, fc1 and fc2.
  - Paged AdamW optimizer.
  - Decoding with temperature 0.7 and top-p 0.9.
  - About 5 GB of memory, compared with about 13 GB for full fine-tuning.
- Augmentation: WordNet synonyms with the same POS replace at most 20% of the content words in a response, and anchor words are never touched. This raises the training set to at most 21,812 examples.
- Epoch ablation:
  - At epoch 2, only 62% of outputs carried the JOKE: marker.
  - From epoch 4 on, adherence was 100%, with dev PPL 1.65 at epoch 4.
  - By epoch 10, the model mostly repeated training jokes.
- Error analysis of 500 dev outputs from the submitted checkpoint:
  - 50% memorized or stock puns on frequent topics such as coffee and the office.
  - 30% broke the required format.
  - 20% showed semantic drift, where the tone stays humorous but the link to the input is lost.

## Interesting ideas
- Treat format compliance as a metric of its own, separate from loss, and stop training where it saturates while dev and train curves are still close. This catches both undertrained and memorizing checkpoints.
- During augmentation, lock the tokens that carry the task constraint and perturb only the rest, with a cap on how much may change. The augmented data then cannot accidentally solve or destroy the task signal.
- Rewrite inputs between active and passive voice to vary the surface form while keeping the meaning.
- Read unusually low perplexity on generated text as a warning that the model is reproducing training data rather than composing something new.
- The paper calls one failure "semantic drift": the style signal overpowers the content signal from a short input, so the output sounds right but no longer relates to the input.

## For Mova

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

- **I1** (learn): Report the rate of well-formed UD output next to LAS for each checkpoint of gparse, and choose the checkpoint where the dev/train gap first opens.
- **I2** (tokens): When generating silver variants for the tagger or lemmatizer, keep target lemmas and named entities frozen and swap only same-POS synonyms elsewhere.
- **I3** (frames): Run a voice-alternation test on prag; a sentence and its passive counterpart should produce identical who-did-what-to-whom frames, and any mismatch flags a role-assignment bug.
- **I4** (back): When the generator or compressor outputs text that nearly matches a training sentence, compare its graph with the source graph to tell copying apart from faithful regeneration.
- **I5** (back): Parse each generated text again and check that its entities and frames overlap the input meaning graph, rejecting outputs that kept the style but lost the content.

*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: [Fatima Jinnah Women University](https://ror.org/009026n40).

📚 Shelves (6): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works · [💭 Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/) · 1,552 works · [💭 Social Psychology](https://md.research.elamur.ai/free/en/shelves/psychology/social-psychology/) · 306 works

🧭 Topics (1): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works

## Cite

Cite as: Insa Abbas and Sadaf Abdul Rauf. 2026. SLPGFJWUInsa at SemEval-2026 Task 1: Enhancing Linguistic Creativity for English Text-Based Humor. 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.

```bibtex
@inproceedings{abbas2026slpgfjwuinsa,
  title = {{SLPGFJWUInsa at SemEval-2026 Task 1: Enhancing Linguistic Creativity for English Text-Based Humor}},
  author = {Insa Abbas and Sadaf Abdul Rauf},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.231},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.231}
}
```

- [← previous SlovKE: A Large-Scale Dataset and LLM Evaluation for Slovak Keyphrase Extraction](https://md.research.elamur.ai/free/en/papers/stevanak-2026-slovke-large-scale-dataset-llm/)
- [next → Small Agents, Big Gains: Journey-Aware and Critic-Guided Simulation for Long-Horizon Shopping Dialogues](https://md.research.elamur.ai/free/en/papers/ping-2026-small-agents-big-gains-journey/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/43/)

## References

13 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Kim Binsted. 1996. Machine Humour: An Implemented Model of Puns. University of Edinburgh. [find](https://scholar.google.com/scholar?q=Kim+Binsted%2E+1996%2E+Machine+Humour%3A+An+Implemented+Model+of+Puns%2E+University+of+Edinburgh%2E)
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: Mwahaha, models write automatic humor and humans annotate. In Proceedings of the 20th International Workshop on Semantic Evaluation (SemEval-2026). [find](https://scholar.google.com/scholar?q=Santiago+Castro%2C+Luis+Chiruzzo%2C+Santiago+G%C3%B3ngora%2C+Salar+Rahili%2C+Naihao+Deng%2C+Ignacio+Sastre%2C+Victoria+Amoroso%2C+Guillermo+Rey%2C+Aiala+Ros%C3%A1%2C+Guillermo+Moncecchi%2C+J%2E+A%2E+Meaney%2C+Juan+Jos%C3%A9+Prada%2C+and+Rada+M)
3. Peng-Yu Chen and Von-Wun Soo. 2020. Humor in word embeddings: Cockamamie gobbledegook for nincompoops. In Proceedings of the 37th International Conference on Machine Learning (ICML). [find](https://scholar.google.com/scholar?q=Peng%2DYu+Chen+and+Von%2DWun+Soo%2E+2020%2E+Humor+in+word+embeddings%3A+Cockamamie+gobbledegook+for+nincompoops%2E+In+Proceedings+of+the+37th+International+Conference+on+Machine+Learning+%28ICML%29%2E)
4. Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. 2023. [Qlora: Efficient finetuning of quantized llms](https://md.research.elamur.ai/free/en/papers/dettmers-2023-qlora-efficient-finetuning-quantized-llms/). In Advances in Neural Information Processing Systems (NeurIPS).
5. He He, Zhiyu Niu, Zhe Geng, and Xiaojun Fan. 2019. Pun generation with surprise. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP). [↗](https://doi.org/10.48550/arxiv.1904.06828)
6. Jack Hessel, Ana Marasović, Jena D. Hwang, Lillian Lee, Hal Daumé III, and Yejin Choi. 2023. Androids laugh at ambiguous anecdotes: Classifying humor in llms. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (ACL). Step Description 1 2 3 4 5 6 7 [find](https://scholar.google.com/scholar?q=Jack+Hessel%2C+Ana+Marasovic%CC%81%2C+Jena+D%2E+Hwang%2C+Lillian+Lee%2C+Hal+Daum%C3%A9+III%2C+and+Yejin+Choi%2E+2023%2E+Androids+laugh+at+ambiguous+anecdotes%3A+Classifying+humor+in+llms%2E+In+Proceedings+of+the+61st+Annual+Meetin)
7. Nabil Hossain, John Krumm, Sayyed Sajadi, and Henry Kautz. 2019. Puns you can rely on: A systematic study of pun generation. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL). *Table 9: Dataset reproduction steps.* Nabil Hossain, John Krumm, Sayyed Sajadi, and Henry Kautz. 2020. Stimulating creativity with tunable language models. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 3432–3443. [find](https://scholar.google.com/scholar?q=Nabil+Hossain%2C+John+Krumm%2C+Sayyed+Sajadi%2C+and+Henry+Kautz%2E+2019%2E+Puns+you+can+rely+on%3A+A+systematic+study+of+pun+generation%2E+In+Proceedings+of+the+57th+Annual+Meeting+of+the+Association+for+Computatio)
8. 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](https://md.research.elamur.ai/free/en/papers/hu-2021-lora-low-rank-adaptation-large/). In International Conference on Learning Representations (ICLR).
9. Mojan Javaheripi, Sébastien Bubeck, Marah Abdin, and 1 others. 2023. [Phi-2: The surprising power of small language models](https://md.research.elamur.ai/free/en/papers/lin-2023-vila-pre-training-visual-language/). arXiv preprint arXiv:2312.07533. [↗](https://arxiv.org/abs/2312.07533)
10. Sophie Jentzsch and Kristian Kersting. 2023. Chatgpt is getting witty: Large language models as a platform for humor generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP). [find](https://scholar.google.com/scholar?q=Sophie+Jentzsch+and+Kristian+Kersting%2E+2023%2E+Chatgpt+is+getting+witty%3A+Large+language+models+as+a+platform+for+humor+generation%2E+In+Proceedings+of+the+2023+Conference+on+Empirical+Methods+in+Natural+L)
11. S. Mangrulkar and 1 others. 2023. Parameter-efficient fine-tuning of large language models. Hugging Face Documentation / PEFT Library. [find](https://scholar.google.com/scholar?q=S%2E+Mangrulkar+and+1+others%2E+2023%2E+Parameter%2Defficient+fine%2Dtuning+of+large+language+models%2E+Hugging+Face+Documentation+%2F+PEFT+Library%2E)
12. George A. Miller. 1995. [Wordnet: A lexical database for english](https://md.research.elamur.ai/free/en/papers/miller-1994-wordnet/). In Communications of the ACM, volume 38, pages 39–41.
13. Tristan Miller, Christian Hempelmann, and Iryna Gurevych. 2017. [Semeval-2017 task 7: Detection and interpretation of english puns](https://md.research.elamur.ai/free/en/papers/miller-2017-semeval-2017-task-7-detection/). In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 58–68. \### Instruction: Generate a humorous response.

*Updated 1 November 2026*