Understanding Laughter toward Authentic Human-Machine Interaction

Abstract

Humans interact with each other with social signals that consist of verbal and non- verbal signals. Therefore, it is important for the model to understand these signals for developing genuine human-machine interaction systems. Despite significant progress in large language models (LLMs) and chatbot systems for processing verbal cues, understanding non-verbal signals presents a notable challenge. Among social signals, laughter is one of the distinctive expressions that occurs during social interactions between humans. In this thesis, I tackle a new challenge for machines to understand the rationale behind laughter in video, Video Laugh Reasoning. I introduce this new task to explain why people laugh in a particular video and a dataset for this task. The proposed dataset, SMILE, comprises video clips and language descriptions of why people laugh. I propose a baseline by leveraging the reasoning capacity of large language models (LLMs) with textual video representation. I further investigate the scalability of the baseline by probing other video understanding tasks and in-the-wild videos. The ability to infer the reason for laughter can be utilized to create models that understand context and laugh appropriately. Overall, this thesis makes significant strides in understanding laughter in video. By doing so, it lays a foundational stone for equipping machines with the ability to understand human-like social signals, marking a crucial step towards more empathetic and authentic human-machine interaction.