In this work, we participate in SemEval-2026 Task 2, which focuses on predicting continuous valence and arousal trajectories from longitudinal ecological essays. To model fine-grained emotional dynamics, we explore three approaches: (1) hierarchical encoder-based models to capture contextual emotional patterns, (2) a lexicon-based pipeline with linguistic rules and a dual-level calibration mechanismfor personalized estimation, and (3) a hybrid framework that integrates lexical emotional signals into neural encoders. Experiments on the official dataset, evaluated using Pearson correlation (r) and MAE, show consistent improvements over baseline methods, highlighting the complementary strengths of neural representations and calibrated lexical features.