Classificação de trocadilhos em português com BERTimbau large: desafios e resultados

Abstract

This paper presents a study on the automatic detection of puns in Portuguese, a significant challenge in the field of Natural Language Processing (NLP) due to semantic ambiguity and context dependence. For the analysis, we used the Puntuguese corpus, a unique collection of 3,990 short texts, equally divided between puns and non-puns, from various sources such as blogs, social media, and YouTube videos. The BERTimbau Large model, specialized in variations of Brazilian and European Portuguese, was applied to analyze the sentences. The work also involved preprocessing techniques and the extraction of linguistic patterns using regular expressions, which aided in data analysis. The task of identifying puns is challenging due to the particularities and subtleties of this type of humor, requiring a deep understanding of linguistic structures. This study offers an important contribution to the advancement of verbal humor detection in Portuguese and opens doors for future developments in the field