Humor Mechanics: Advancing Humor Generation with Multistep Reasoning
Alexey Tikhonov, Pavel Shtykovskiy
TL;DR
The paper investigates generating one-liner humor with multi-step reasoning by learning a data-driven humor-generation policy from jokes. It infers a policy $P_{humor_policy}$ from a seed joke dataset, decomposes jokes with GPT-4 to capture building blocks, and uses iterative associations to guide a multistep generation pipeline, formalized as $\pi_{humor} = LLM(P_{humor_policy} + P_{seed} + P_{associations})$. Human evaluation on ScaleAI shows the full pipeline improves novelty and maintains or increases understandability and funniness relative to baselines, while producing more original jokes than zero-shot GPT-4 and Reddit-based jokes. The work highlights the importance of brainstorming associations and culture-aware policy prompts for creative AI, and discusses ethical considerations and avenues for future work in broader humor and AI creativity tasks.
Abstract
In this paper, we explore the generation of one-liner jokes through multi-step reasoning. Our work involved reconstructing the process behind creating humorous one-liners and developing a working prototype for humor generation. We conducted comprehensive experiments with human participants to evaluate our approach, comparing it with human-created jokes, zero-shot GPT-4 generated humor, and other baselines. The evaluation focused on the quality of humor produced, using human labeling as a benchmark. Our findings demonstrate that the multi-step reasoning approach consistently improves the quality of generated humor. We present the results and share the datasets used in our experiments, offering insights into enhancing humor generation with artificial intelligence.