Exploring Humor Detection in Bengali: A Case Study Leveraging Transformer Models

The widespread use of online social media platforms has led to a plenty of user-generated data, where users express their opinions in real-time. Humor plays a crucial role in facilitating engaging conversations, as it has the ability to enliven dull conversations. Understanding humor and its linguistic complexities presents a challenge for computers, particularly in low-resource languages like Bengali. While humor detection has been extensively studied in resource-rich languages due to the availability of large data, it remains challenging for languages with limited resources. In this study, we addressed this challenge by focusing on low-resource languages, such as Bengali, through data collection and annotation. We conducted automatic humor detection experiments using various transformer models, including Multilingual BERT, XLM-RoBERTa, DistilBERT and BanglaBERT Base to establish a baseline detection system. Our findings revealed that DistilBERT outperformed other models, achieving an F1-score of 0.897 with 89.4% accuracy.