Let's be Humorous: Knowledge Enhanced Humor Generation
Level 5 - mechanism / opinion, no new human data
By design analogy for computational/NLP model development and benchmark paper without clinical data.
OpenAlex W3022525654 · doi:10.48550/arxiv.2004.13317
What was done
The authors proposed an end-to-end framework to generate humor punchlines from given set-ups by incorporating relevant background knowledge. They also created a new humor-knowledge dataset and evaluated their model against several baseline methods.
What was found
The abstract reports no numerical values, statistical comparisons, or benchmark scores. It states qualitatively that the framework successfully leveraged knowledge to generate fluent and funny punchlines, outperforming baseline models.
Why it matters
This approach explores moving beyond template-based or phrase-replacement methods by fusing external knowledge into freer-form punchline generation models.
Limits
The abstract reports no quantitative evaluation metrics, dataset size, or details on how fluency and humor were judged (such as human evaluators vs. automated metrics). As a preprint, it has not undergone formal peer review.
Cited by
- context Current transformer-based AI models lack the structural mechanism to start with a punchline and work backwards to generate a coherent setup leading up to it.