Year: 2026
Type: conference-paper
Abstract: SemEval-2026 Task 1 (MWAHAHA) includes a subtask focused on text-based humor generation in English, where systems are required to produce humorous content from minimal textual input such as headlines or word pairs.Humor generation presents unique computational challenges due to its reliance on pragmatic inference, expectation violation, and stylistic nuance.This paper describes a prompt-based generation framework designed to produce concise, dry, and satirical headlines that follow strict stylistic and safety constraints.The proposed approach operates in a zero-shot setting using a GPT-4-class large language model without task-specific fine-tuning.Structured prompt engineering is employed to enforce constraints including single-sentence output, 8-12 word length, deadpan tone, and subtle expectation subversion.Deterministic decoding ensures consistency and replicability across runs.Output validation guarantees strict adherence to the official submission requirements.Experimental analysis includes prompt ablation comparisons and qualitative evaluation of generated outputs.Results indicate that carefully engineered constraints significantly improve stylistic alignment compared to unconstrained prompting.The system demonstrates that controlled humor generation can be achieved through structured prompting without supervised training, highlighting the effectiveness of instruction-based large language model utilization for creative NLP tasks.
Authors: Hemeshkumar Parthiban, Priyadharsini R
Institution: Rajalakshmi Engineering College
Language: English
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166850339