EmoTracker—A New Framework for Modeling and Forecasting Diachronic Emotion Dynamics

Abstract

Existing computational approaches to diachronic semantics and emotion analysis typically study word meaning change and emotional evolution separately, limiting our understanding of how emotions evolve in proportion to the sense level. To bridge this gap, we propose EmoTracker, a novel framework that integrates diachronic sense modeling with Valence-Arousal-Dominance (VAD) emotion tracking to model and predict temporal emotion-sense trajectories. Our contribution is threefold. First, we develop a method for constructing temporal emotion datasets by integrating diachronic sense data with three different VAD lexicons. Second, we implement an LSTM architecture with attention mechanisms and momentum-based features to forecast emotional trajectories over time. Third, we provide interactive 3D visualizations to explore emotion dynamics over time, and 4D visualizations to capture the diachronic joint evolution of emotions and senses in the VAD space. Our evaluation shows that, among the selected lexicons, NRC-VAD is the most suitable for temporal modeling, though it also reveals the challenges in modeling dominance across lexicons. EmoTracker bridges diachronic semantics and emotion analysis, providing a comprehensive framework for computational humanities research.