Modeling Continuous Emotions in Text Data using IEMOCAP Database

Predicting emotional state and specifically valence (positivity or negativity), arousal (activation or intensity), and dominance (control or power) from text data remains a challenge due to the complexity and context dependency of human emotions. Recent advancements in deep learning, especially Transformer models like BERT, have significantly enhanced text analysis capabilities by capturing complex semantic relationships within text. This study explores the performance of various deep learning architectures in predicting emotional states using a speaker-dependent(SD) scenario for data splitting, where both training and testing sets contain samples from the same speakers. We compare a baseline Long Short Term Memory (LSTM) model, an LSTM model with GloVe word embeddings, and a pretrained BERT model, all evaluated on the IEMOCAP dataset. Our Research shows that the BERT model achieves the highest Concordance Correlation Coefficient (CCC) scores 0.724 for valence, $\mathbf{0. 423}$ for arousal, and $\mathbf{0. 488}$ for dominance resulting in an average CCC score of 0.545, outperforming both the baseline LSTM and the LSTM+GloVe models. These results highlight Transformer’s ability especially BERT to capture continuous emotions in textual data.