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

2026-06-27 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

A study investigates the generalizability of humor detection in machines, addressing whether models trained on specific humor types can transfer competence to novel, unseen forms. Given the continuous emergence of new online humor, such as memes and "AI fails," the research explores if large language models (LLMs) can capture deeper, transferable mechanisms. Researchers conducted transfer learning experiments using four distinct humor datasets, varying training diversity from one to three datasets before testing on a novel one. Results indicate that models achieve up to 75% accuracy on binary unseen datasets, demonstrating some transfer capability. Crucially, training on diverse sources enhanced transferability by 1.88-4.05% with negligible impact on in-domain performance. Interestingly, the "Dad Jokes" dataset proved most effective for enabling transfer, yet it was also the most challenging target for transfer. The associated data and code have been released.

Key takeaway

For NLP Engineers developing humor detection systems, you should prioritize training models on diverse humor datasets to enhance generalization to new, emerging humor types like memes or "AI fails." Your approach should leverage varied sources, as this improves transferability by 1.88-4.05% with minimal in-domain performance loss. Consider incorporating "Dad Jokes" into your training regimen, as it proved a strong enabler for transfer, despite being a challenging target itself.

Key insights

Models can generalize humor detection across diverse types, improving with varied training data.

Principles

• Competence on specific humor tasks can transfer to novel types.
• Diverse training sources improve humor detection transferability.
• Some humor types are better enablers for transfer.

Method

Conduct transfer learning experiments across four datasets, varying training diversity (1-3 datasets) and testing on a novel, unseen humor type.

In practice

• Train humor detection models on diverse datasets for better generalization.
• Utilize "Dad Jokes" as a foundational dataset for humor transfer learning.

Topics

• Humor Detection
• Transfer Learning
• Computational Humor
• Large Language Models
• Dataset Diversity
• Dad Jokes

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Memes-as-Replies: Can Models Select Humorous Manga Panel Responses? — LLMs show promise in understanding social cues for humor, but struggle with visual information and subtle wit in meme reply selection.
• Humor Style Drives Laughter, Topic Shapes Acceptability: Evaluating Bilingual Personal and Political Robot-Delivered AI Jokes — Humor style dictates funniness, while joke topic determines acceptability for robot-delivered AI jokes.
• Re-defining Humor Data Objects for AI Humor Research — AI humor research improves by defining humor as a social interaction with context and explanations, not just a binary state.
• Identification of fake news in Portuguese: a look at the generalization of models — Language models for fake news detection show significant performance degradation in cross-generalization scenarios.
• Adapting Reinforcement Learning with Chain-of-Thought Supervision for Explainable Detection of Hateful and Propagandistic Memes — Reinforcement learning with CoT supervision significantly improves MLLM performance and explainability for hateful meme detection.
• Token-Level Pun Location Using Multi-Layer BERT with Mixture of Experts — A two-stage BERT-MoE model effectively locates puns in Portuguese by specializing expert components.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.