From Humor Categories to Funniness Prediction

Sebastião Miguel Gomes Pereira Lemos Caldas

Key information

Authors:

Sebastião Miguel Gomes Pereira Lemos Caldas (Sebastião Miguel Gomes Pereira Lemos Caldas)

Supervisors:

Helena Sofia Andrade Nunes Pereira Pinto (Helena Sofia Andrade Nunes Pereira Pinto); Maria Luísa Torres Ribeiro Marques da Silva Coheur (Maria Luísa Torres Ribeiro Marques da Silva Coheur)

Published in

November 13, 2024

Abstract

Humor is a multifaceted phenomenon encompassing diverse themes and styles, with quality perceptions varying widely across individuals. Prior research has tried either to categorize humor through taxonomies or to investigate predictive models for individual responses to specific humorous media, but research handling both is either rare or non-existent. This study aims to investigate the predictive value of humor type labels on individual ratings of humorous content, assessing whether identifying humor types within textual humor can enhance prediction of user-specific enjoyment. I used the Jester dataset, which includes user-assigned funniness ratings, as a starting point, extending it with a humor taxonomy, and labeling each joke according to distinct humor techniques. Two experiments are conducted in this work. First, jokes and users are clustered, and the relationship between these clusters and humor techniques is examined. Findings indicate a lack of significant association between humor type and perceived funniness across both joke and user clusters. In the second experiment, predictive models (based on joke embeddings inputs from Large Language Models) for funniness ratings are trained, testing the impact of incorporating humor type labels as input features. Results similarly suggest that humor type labels do not significantly enhance model performance. This work provides an initial exploration into the relationship between humor taxonomies and subjective funniness, contributing with some insights for future studies in humor prediction.