DUTH at SemEval-2026 Task 1: Prompt-Based Zero-Shot Large Language Models for Constrained Multilingual Humor Generation

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

Summary

DUTH's system for SemEval-2026 Task A addresses the challenging problem of constrained multilingual humor generation across English, Spanish, and Chinese. This approach employs instruction-tuned large language models in a zero-shot configuration, integrating prompt engineering, controlled decoding, and lightweight post-generation validation. The validation step ensures constraint satisfaction and language consistency in the generated short, humorous outputs. Researchers evaluated various model families and parameter scales, including Qwen and Mistral models. Human evaluation results consistently demonstrated that larger models achieved superior performance compared to smaller ones, with the Qwen2.5-14BInstruct model exhibiting the strongest overall results. Error analysis identified persistent issues such as violations of lexical constraints and interference between languages.

Key takeaway

For NLP Engineers developing multilingual content generation systems, you should consider zero-shot instruction-tuned LLMs for humor tasks. Your approach should integrate prompt engineering with controlled decoding and post-generation validation to enforce specific constraints. Prioritize larger models like Qwen2.5-14BInstruct for superior results. However, be prepared to address persistent challenges such as lexical constraint violations and cross-lingual interference in your outputs.

Key insights

Zero-shot instruction-tuned LLMs can generate constrained multilingual humor with specific engineering.

Principles

• Larger LLMs consistently outperform smaller ones for humor generation.
• Humor generation is inherently subjective and culturally dependent.

Method

Apply prompt engineering, controlled decoding, and post-generation validation with instruction-tuned LLMs in a zero-shot setup.

In practice

• Consider Qwen2.5-14BInstruct for multilingual humor tasks.
• Focus on mitigating lexical constraint violations and cross-lingual interference.

Topics

• Humor Generation
• Multilingual NLP
• Zero-Shot Learning
• Large Language Models
• Prompt Engineering
• SemEval-2026

Best for: Research Scientist, AI Scientist, NLP Engineer, Prompt Engineer

Related on AIssential

• Supervision versus Demonstration-Based In-Context Learning for Multiword Expression Classification — Prompt sensitivity and demonstration quality critically impact LLM performance in metalinguistic classification tasks.
• 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.
• Better Literary Translation: A Multi-Aspect Data Generation and LLM Training Approach — Multi-aspect data generation and iterative refinement significantly enhance LLM literary translation.
• As Easy as Rocket Science: Assessing the Ability of Large Language Models to Interpret Negation in Figurative Language — LLMs struggle with combined negation and figurative language, with prompt style significantly impacting interpretation accuracy.
• Zero-Shot to Full-Resource: Cross-lingual Transfer Strategies for Aspect-Based Sentiment Analysis — Cross-lingual transfer strategies for ABSA vary in effectiveness based on model architecture and resource availability.
• BanglaBERT vs. Frontier LLMs: Diagnosing Zero-Shot Collapse in Bangla NLP — Frontier LLMs experience zero-shot collapse on low-resource Bangla NLP tasks, necessitating fine-tuned models or few-shot scaling for nuanced sentiment.

Open in AIssential →

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