Thai Humor Generation by Small Language Models

Despite the impressive capabilities of generative AI across multiple languages, generating humor that aligns with Thai cultural and linguistic nuances remains a significant challenge. Thai humor often relies on context, wordplay, and socio-cultural references, making it difficult for generic models to produce authentic jokes. This paper presents a focused approach to address this limitation by fine-tuning small language models (SLMs) on high-quality, non-synthetic Thai humor datasets. Llama-3.2-3B model was leveraged and Low-Rank Adaptation (LoRA) was employed for efficient parameter tuning, ensuring computational efficiency suitable for low-resource settings. Our work highlights humor as a critical benchmark for evaluating AI’s understanding of language semantics and cultural context. A comprehensive evaluation was conducted with Thai participants to ensure the generated humor resonates with real-world cultural expectations.

Recommended citation: Poobanchean, P., Chaithong, S., Muansuwan, N., Limseelo, C., Sirinaovakul, B., & Suwannahong, K. (2025). Thai Humor Generation by Small Language Models. King Mongkut’s University of Technology Thonburi.
Download Paper | Download Slides