Toward a Multi-dimensional Humor Dataset for Social Robots

Expressing humor in social interactions presents a significant challenge for humans due to its intricate linguistic nature. This complexity is further magnified when teaching robots to express humor appropriately. Among the various expressions of humor, jokes are one of the most commonly used. Therefore, a well-annotated joke dataset holds significant promise in enhancing a robot's ability to express humor effectively. This paper introduces a dataset comprising over two thousand jokes, with the aim of providing rich material and multidimensional selection criteria for the humor expression of the robot. The creation of this joke dataset involved a collaborative effort among robot experts studying HRI, psychologists with rich humor research experience, and GPT-3. The annotation process primarily concentrated on four dimensions within the dataset: the humor style of jokes, semantic words (aligned with semantic gestures), keywords, and ratings of joke funniness. We additionally outline several prospective applications of this dataset. We introduce a BERT-based neural network model trained on the dataset with semantic word labels. This model aims to empower robots to choose suitable semantic words from jokes and articulate them alongside corresponding semantic gestures. Moreover, we offer suggestions for utilizing jokes from this dataset to facilitate the adaptive expression of humor by social robots. These endeavors will further enhance the multi-modal humor expression capability of social robots.