📄 JCT at SemEval-2026 Task 1: Let the Best Joke Win - A Generate - and-Rank Approach to Constrained Humor

Summary

The paper targets English constrained joke writing in SemEval-2026 Task 1, Subtask A. Systems must produce a short joke that either contains two given words or relates to a news headline. Because LLM humor quality is uneven, the JCT team recasts the task as picking the best joke from many candidates rather than trusting a single generation.

For every input, gpt-5.1 drafts jokes using three prompting modes (zero-shot, few-shot and chain-of-thought-guided) crossed with ten humor styles. Candidates that break the constraint are filtered out or regenerated. The survivors are ranked by a hand-weighted mix of five automatic proxies: incongruity, emotion-based humor potential, irony, fluency and novelty. The top-scoring joke is submitted, and no fine-tuning is required.

The analysis looks at which candidates win rather than at human funniness ratings:

• Few-shot and structured prompts are chosen far more often than zero-shot ones.
• Observational humor is the most frequently selected style.

The work shows a cheap, modular way to make LLM humor more reliable. However, the authors acknowledge that the proxies are not validated against human judgments.

Key points

• Candidate generation: 3 prompting strategies × 10 styles yield 30 candidates per constraint. The ten styles cover dark humor, wholesome humor, wordplay, irony, the absurd, incongruity, self-deprecation, observational comedy, logical twists and social humor.
• Constraint checks:
  • Word constraints require both words to appear exactly.
  • Headline constraints require embedding cosine similarity of at least 0.6.
  • Failing candidates are retried up to 4 times.
• Scoring models and weights:
  • Incongruity: all-mpnet-base-v2 setup/punchline dissimilarity, weight 0.35.
  • Humor potential: roberta-base-go_emotions amusement/surprise scores, weight 0.25.
  • Novelty: all-MiniLM-L6-v2 compared against the Kaggle short-jokes corpus, weight 0.15.
  • Fluency: GPT-2 perplexity, weight 0.15.
  • Irony: cardiffnlp twitter-roberta-base-irony, weight 0.10.
• Results and limitations: Winning jokes come mostly from few-shot prompts, then chain-of-thought prompts. Observational, irony and self-deprecating styles dominate, while wordplay, incongruity and wholesome jokes are rarely chosen. The weights were set manually, and no human evaluation was performed.
• Code: The code is released at github.com/batyaSchechter/Humor_generation.