Research map / Large language models / Language, safety and interpretability

Talk Funny! A Large-Scale Humor Response Dataset with Chain-of-Humor Interpretation

AAAI 2024 · Large language models › Language, safety and interpretability · approach: Data and evaluation methodology · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Builds TalkFunny, a 4,116-pair Chinese humor response dataset with chain-of-humor and humor mind map annotations plus two auxiliary humor tasks, and fine-tunes PLMs to show these annotations improve humorous response generation.

Problem

Pre-trained language models generate humorous dialogue responses poorly, largely due to a shortage of humor datasets and the tacit knowledge that humor generation requires.

Approach

Annotate each humor context-response pair with an explicit chain-of-humor explanation and a humor mind map (knowledge triplets), and jointly train response generation with humor sentiment-style classification and humor rewriting as auxiliary tasks.

Key results

• Multitask-based annotation-enhanced CPT-large raises BARTScore from -6.3 (original) to -4.9 and Dist-1 from 44.6 to 50.9 on TalkFunny.
• GPT-3.5 with a humor prompt instruction reaches a manual humor degree of 91.1 versus 71.2 without the prompt, while fine-tuned PLMs with original inputs score around 48-49.
• Ablation shows chain-of-humor contributes more than humor mind maps, and humor rewriting gives the largest gain among auxiliary tasks.

Measured effects reported in the paper

Surprising findings

• Fine-tuned small PLMs without annotations perform similarly across model sizes, and the paper reads this as the dataset being robust rather than model capacity mattering.
• GPT-3.5 without prompt instructions performs poorly at humor generation, while adding a short humor prompt yields large gains.

Limitations

• Some generated responses are not fluent or are hard for humans to understand, and the chain-of-humor and mind maps are hard for PLMs to learn.
• Dataset is Chinese-only and small (4,116 pairs), with annotations partly produced with GPT-3.5 and proofread by humans.
• Human evaluation covers only 1,000 generated responses per PLM with three volunteers, and the humor degree metric uses ad hoc weights.

Experimental setup

Datasets: TalkFunny

Benchmarks: TalkFunny

Metrics: Dist-1 Dist-2 Greedy Match Embedding Average BARTScore unexpectedness personification sentiment correlation informativeness coherence

Baselines: T5 BART CPT GPT-3 GPT-3.5

Builds on: T5 BART CPT BARTScore

Closest papers by meaning

• HUMORCHAIN: Theory-Guided Multi-Stage Reasoning for Interpretable Multimodal Humor Generation (CVPR 2026 · similarity 0.88)
• Innovative Thinking, Infinite Humor: Humor Research of Large Language Models through Structured Thought Leaps (ICLR 2025 · similarity 0.86)
• Humor in AI: Massive Scale Crowd-Sourced Preferences and Benchmarks for Cartoon Captioning (NeurIPS 2024 · similarity 0.85)
• From Perception to Punchline: Empowering VLM with the Art of In-the-wild Memes (NeurIPS 2026 · similarity 0.85)
• Cracking the Code of Juxtaposition: Can AI Models Understand the Humorous Contradictions (NeurIPS 2024 · similarity 0.84)
• Let's Think Outside the Box: Exploring Leap-of-Thought in Large Language Models with Creative Humor Generation (CVPR 2024 · similarity 0.83)