Performance and Complexity Analysis of Neural Architectures for Emotion Classification Using Word Embeddings

The paper evaluates five deep learning architectures RNN, LSTM, GRU, Seq2Seq (LSTM Encoder), and Transformer Encoder for multi-class emotion classification across six fine-grained emotional categories (sadness, joy, love, anger, fear, and surprise). To capture underlying semantic representations, the study leverages pre-trained Word2Vec and GloVe word embeddings, analyzing both baseline embedding structures (via t-SNE and UMAP visualizations) and model performance following task-specific fine-tuning.