CUETClashing at SemEval-2026 Task 1: Multilingual Joke Generation Under Lexical and Topical Constraints Using Small Instruction-Tuned LLMs

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

Summary

CUETClashing's system for SemEval-2026 Task 1, "MWAHAHA - Models Write Automatic Humor And Humans Annotate," addresses the challenging task of multilingual joke generation. This system creates single-sentence jokes in English, Spanish, and Chinese, adhering to two specific rules: incorporating designated words and relating to a provided news headline. The approach utilizes small instruction-tuned language models, specifically Qwen2.5-3B-Instruct for English and Chinese, and Salamandra-2B-Instruct for Spanish. It integrates language-specific prompts, a special sampling technique for outputs, and a robust post-generation cleaning process. Notably, the system achieves rule adherence in all three languages without requiring additional task-specific training, demonstrating the effectiveness of simple prompt design and smaller, instruction-tuned models for efficient, cross-lingual humorous text generation.

Key takeaway

For NLP Engineers developing multilingual content generation systems, this work suggests that you can achieve complex, constrained text generation without extensive task-specific fine-tuning. If your goal is to produce creative outputs like humor across languages, consider using smaller, instruction-tuned LLMs such as Qwen2.5-3B-Instruct or Salamandra-2B-Instruct. Focus on designing effective, language-specific prompts and robust post-generation cleaning to meet specific lexical and topical constraints efficiently.

Key insights

Small instruction-tuned LLMs, combined with careful prompting and post-processing, can effectively generate constrained, multilingual humorous text without task-specific training.

Principles

• Humor generation requires creativity, cultural understanding, and rule adherence.
• Small instruction-tuned LLMs can be highly effective.
• Task-specific training is not always necessary for complex NLG.

Method

The method employs instruction-tuned LLMs (Qwen2.5-3B-Instruct, Salamandra-2B-Instruct) with language-specific prompts, special output sampling, and a strong post-generation cleaning process for constrained multilingual joke creation.

In practice

• Use small instruction-tuned LLMs for constrained NLG tasks.
• Implement language-specific prompts for multilingual outputs.
• Incorporate post-generation cleaning for quality control.

Topics

• Multilingual NLP
• Joke Generation
• Instruction-Tuned LLMs
• Qwen2.5-3B-Instruct
• Salamandra-2B-Instruct
• SemEval-2026 Task 1
• Natural Language Generation

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• English is Not All You Need: Systematically Exploring the Role of Multilinguality in LLM Post-Training — Multilingual post-training significantly improves LLM performance across languages and tasks, outperforming English-only approaches.
• Automatic Generation of Highlights for Academic Paper Via Prompt-based Learning — Prompt-based learning with LLMs can generate academic paper highlights effectively without extensive labeled data.
• Memes-as-Replies: Can Models Select Humorous Manga Panel Responses? — LLMs show promise in understanding social cues for humor, but struggle with visual information and subtle wit in meme reply selection.
• Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs with Controllable Difficulty and Safety — Fine-tuning compact 8B LLMs can generate children's stories with controllable difficulty and safety, outperforming larger models.
• Constrained Semantic Decompression in LLMs through Persian Proverb-Conditioned Story Generation — LLMs struggle with "constrained semantic decompression," failing to translate abstract cultural knowledge like proverbs into faithful narratives.
• Low-Latency Real-Time Audio Game Commentary System via LLM-Based Parallel Text Generation — Parallel text generation and utterance buffering significantly reduce latency and improve rhythm in real-time audio commentary systems.

Open in AIssential →

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