CSIRO-LT at SemEval-2026 Task 2: In-the-Wild Valence and Arousal Forecasting on Ecological Text Time Series

AbstractPredicting emotional valence and arousal in text is challenging due to the continuous, dynamic, and context-dependent nature of emotions. The SemEval 2026 Task 2: Predicting Variation in Emotional Valence and Arousal over Time from Ecological Essays shared task investigates longitudinal affect prediction from real-world personal essays, including forecasting short-term state and longer-term dispositional changes. We compare Pre-trained Language Models (PLMs) and Large Language Models (LLMs) for these subtasks, examining different input representations and feature formulations. We show that sentiment-aware PLMs are most effective for continuous valence and arousal prediction, and LLMs are effective for short-term state forecasting. Modelling dispositional changes remains challenging, and none of our neural approaches surpass simple a historical baseline approach in this setting.