Text Rewriting with Transformers for Humor Generation in Portuguese

André Moreira, Marcio Lima Inácio, Hugo Gonçalo Oliveira

Abstract: Humor Generation has been tackled by Artificial Intelligence for a long time, but remains a challenge, in particular for less-resourced languages. We explore Humor Generation in Portuguese through the task of rewriting a text to make it funny, taking advantage of data available for Portuguese and of transformers, which were: (i) fine-tuned for the task (T5); (ii) used as the mutation operator and fitness function (BERT) of a Genetic Algorithm that evolves text; (iii) prompted for the task (via instruction-tuned LLMs). Given the subjectivity of Humor, funniness of the produced text and its relation to the input was scored by human judges. Results reveal that the only approaches that produce potentially funny text most of the time are those based on prompting LLMs, which suggests that the complexity of humor is better targeted with larger models, even if not specifically trained for the task.

External IDs:dblp:conf/epia/MoreiraIO25