Cherish at SemEval-2026 Task 2: Enhancing RoBERTa-Based Models for Emotional Valence and Arousal Prediction in Ecological Essays with Personalized PLoRA and Temporal Embeddings

AbstractThis paper describes the system developed by Team Cherish for SemEval-2026 Task 2: Predicting Variation in Emotional Valence and Arousal over Time from Ecological Essays. Our approach models emotional dynamics in user-generated text by incorporating both personalization and temporal information into a transformer-based architecture. We use RoBERTa-large as the backbone encoder and enhance it with PLoRA and a temporal embedding module. Cherish’s model architecture is designed to maintain general semantic knowledge while subtly adapting to individual users and emotional shifts over varying temporal gaps. Our system achieved 13th place out of 29 teams in Subtask 1, obtaining a Pearson’s r composite score of 0.596 for valence prediction and 0.505 for arousal prediction. While the team also participated in Subtask 2a, technical issues during inference led to zero variance in predictions, resulting in an undefined (NaN) official correlation score.