DyEmo - Crowdsourcing emotion trajectories: Decoding emotion dynamics from Danmu during naturalistic movie viewing

Emotions are inherently dynamic, yet existing methods fail to capture their fine-grained temporal evolution in naturalistic contexts. Here, we introduce a large language model (LLM)-based framework that decodes high-resolution, multidimensional emotion dynamics from massive-scale crowdsourced Danmu (i.e., bullet-screen comments) during naturalistic movie viewing. Analyzing more than 7 million Danmu comments from over 100 widely viewed full-length movies, we derive continuous ratings across multiple emotion categories, producing emotion trajectories with second level temporal resolution. These trajectories align closely with human annotations and demonstrate robustness across various LLM archi tectures, Danmu sender cohorts, posting-year cohorts, and languages. Leveraging this data, we quantify key dynamical properties—such as inertia, instability, controllability, and self-similarity. Furthermore, we show that the dynamic emotion space is structured around three dimen sions—polarity, complexity, and intensity—with population responses forming a continuous mixed-emotion basin rather than distinct emo tional categories. This scalable and ecologically grounded approach provides a powerful framework for understanding emotion dynamics in naturalistic viewing.