DyEmo: Danmu captures temporal and state-space structure in population-level emotion dynamics

DyEmo (Dynamic Emotion from Danmu) is a dataset for studying population-level emotional expression dynamics from large-scale, time-aligned Danmu comments using large language models (LLMs). The dataset contains Danmu streams, LLM-derived multidimensional emotion trajectories, and human annotations for selected benchmark datasets. It covers: - **102 full-length movies** from Bilibili with large-scale Danmu comments; - **13 bilingual anime episodes** with Chinese and Japanese Danmu for cross-language analysis; - **18 short videos** with human annotations for emotion inference evaluation; - **Forrest Gump** with additional human ratings and vision-language model (VLM) comparison data. DyEmo is designed to support research on naturalistic emotion dynamics, including temporal evolution of emotional expression, narrative affect, cross-language consistency, and computational modeling of population-level affective responses. The dataset provides three major types of information: 1. **Danmu comments** - Raw and processed Danmu text; - Time-aligned comment segments and metadata. 2. **LLM-derived emotion trajectories** - Multidimensional emotion scores inferred from Danmu using LLMs; - Continuous emotion trajectories across movies, episodes, and videos.3. **Human annotations and benchmark data** - Human emotion ratings for selected datasets; - Additional model outputs for benchmarking emotion inference.