Abstract
The current study integrated phone-based Ecological Momentary Assessment (EMA) and Large Language Model (LLM) to investigate whether real-life prediction errors (PEs) influence people’s emotional response, following emotion regulation, and memory sentiment after the initial provocative events. A total of 224 participants submitted their predictions about their chemistry exam grades and uploaded multiple audio recordings describing their feelings after viewing their grades for four exams across a period of 17 weeks. A week later, participants submitted another audio recording to describe the scenario when they viewed their grade. All audio recordings were first transcribed, then a LLM classified all transcriptions from the dense sampling period into Emotion Expression, Emotion Regulation (ER), and Neither categories and for all transcriptions a sentiment analysis was conducted. Results from statistical models supported that PE predicted the audio recording sentiment. PE and sentiment score presented an interaction effect on the amount of regulation observed in transcriptions. A more positive PE predicted a more positive sentiment, and less observed regulation. Furthermore, after regulation, the predicted possibility of observing neutral and mixed valence in audio increased, suggesting observed regulation reduced the extreme emotionality. Moreover, regardless of the valence of PE, as the amount of regulation increased, the sentiment of later recalled memory tended to be more neutral. In addition, students at higher risk of GAD presented a larger regulation effect on both negative and positive PE, indicated as a more neutral AM. In summary, the present study connects PE, emotional response, spontaneous ER in real-life, and even emotional memory through EMA and LLM, delineating a trajectory of emotional cascade. Crucially, our results emphasize the bidirectionality of an adaptive ER system and its connection with anxiety.