SLPGFJWUInsa at SemEval-2026 Task 1: Enhancing Linguistic Creativity for English Text-Based Humor
Year: 2026
Type: conference-paper
Abstract: This paper presents the submission of the Speech and Language Processing Group (SPLG), Fatima Jinnah Women University, for SemEval-2026 Task 1 on constrained humor generation.Our system ranked first in human evaluation with an aggregate score of 1080.The task requires generating humorous text under constraints such as rare word pairs and satirical news headlines.We fine-tune the Phi-2 (2.7B) model using parameter-efficient fine-tuning (PEFT) with QLoRA.A key contribution is a custom dataset of 12,118 instances constructed through a fourstage pipeline combining LLM-based generation (ChatGPT-4o and Gemini 1.5 Pro) with human filtering.We apply WordNet-based synonym replacement and three rule-based transformations to improve diversity and reduce memorization.We provide detailed analysis of training behavior, human evaluation methodology, quantitative baseline comparisons, and failure modes including overfitting and semantic drift.Our results demonstrate that small language models, when carefully fine-tuned and constrained, can produce competitive and creative humor.
Authors: Insa Abbas, Sadaf Abdul Rauf
Institution: Fatima Jinnah Women University
Language: English
Subfield: Social Psychology
Field: Psychology
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166859713