BAHAHA at SemEval-2026 Task 1: Benchmarking-Aware Humor Authoring with Hybrid Assessment and Adaptation

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, short

Summary

The BAHAHA system, developed by Utsav Arora and Andrew Hoblitzell for SemEval-2026 Task 1: MWAHAHA, generates original jokes from news headlines or rare word pairs. This system employs a generate-then-rank pipeline, utilizing multi-style candidate generation through comedian-inspired few-shot prompting. Quality assessment is performed by a smaller model fine-tuned on synthetic rating data from the generation model. BAHAHA produces up to 50 candidates per input across 15 stylistic templates, selecting outputs via a mixed-initiative interface combining automated ranking with human judgment. In a contest with 305 participants and 180 submissions, BAHAHA ranked 2nd on Subtask A Chinese and 5th on Subtasks B1 and B2, generating jokes natively in each language.

Key takeaway

For NLP Engineers developing creative text generation systems, consider adopting a generate-then-rank architecture to enhance output quality and diversity. Your systems can benefit from multi-style candidate generation via few-shot prompting and a smaller, fine-tuned model for initial quality assessment. Integrating human judgment into the final selection process, as demonstrated by BAHAHA's success in SemEval-2026, is crucial for subjective tasks like humor authoring.

Key insights

A hybrid generate-then-rank pipeline effectively creates multi-style humor using few-shot prompting and synthetic data for quality assessment.

Principles

• Few-shot prompting can inspire diverse stylistic generation.
• Synthetic data can train smaller models for quality assessment.
• Hybrid human-automated ranking improves output selection.

Method

Generate up to 50 joke candidates per input across 15 stylistic templates using few-shot prompting, then select outputs via a mixed-initiative interface combining automated ranking with human judgment.

In practice

• Experiment with comedian-inspired few-shot prompts for creative text.
• Fine-tune smaller models on synthetic data for efficient ranking.
• Implement mixed-initiative interfaces for subjective content tasks.

Topics

• Humor Generation
• SemEval
• Natural Language Generation
• Few-Shot Learning
• Ranking Models
• Multilingual NLP

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• I built a public voting benchmark where models have to make memes out of daily news — Memebench evaluates LLM humor and creativity by generating memes from daily news for public A/B voting.
• Part 2: Turning My Writing into a Training Set — A method called "grounded generation" creates personalized fine-tuning datasets by constraining LLMs to source documents, ensuring authentic voice and content.
• CAF-Gen: A Multi-Agent System for Enriching Argumentation Structures — CAF-Gen's multi-agent Creator-Reviewer system enriches shallow argument structures into robust, CAF-compliant models, overcoming single-pass generation instability.
• 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.
• Severity-Aware Curriculum Learning with Multi-Model Response Selection for Medical Text Generation — Severity-aware curriculum learning with multi-model response selection significantly improves medical text generation quality and relevance.
• 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.

Open in AIssential →

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