SMILE-Next: A Multimodal Dataset for Laughter Understanding in LLMs
A new dataset named SMILE-Next has been launched by researchers to facilitate the understanding of laughter in real-life situations. This dataset includes multimodal textual representations and question-answer annotations for three specific tasks: detecting laughter, classifying laughter types, and reasoning about laughter. The goal is to create a large language model (LLM) tailored for laughter comprehension in various contexts. Two main components are introduced: Self-Instruct for laughter, which improves generalization by generating a variety of laughter-focused instructions, and the Mixture-of-Laugh-Experts (MoLE) framework, which presents a specialized expert architecture for tasks. This work aims to deepen the analysis of laughter as a complex social signal. The findings are detailed in a paper available on arXiv (ID: 2605.28084).
Key facts
- SMILE-Next is a dataset for real-world laughter understanding.
- It includes multimodal textual representations and question-answer annotations.
- Three tasks: laughter detection, laughter type classification, and laughter reasoning.
- Aims to develop a laughter-specialized large language model.
- Laughter-specific Self-Instruct synthesizes diverse laughter-centric instructions.
- Mixture-of-Laugh-Experts (MoLE) is a task-specific expert framework.
- Laughter is described as a complex social signal conveying communicative intent.
- The paper is available on arXiv (ID: 2605.28084).
Entities
Institutions
- arXiv