From Statistical Features to Contextual Embeddings: A Multi-Paradigm Evaluation of Emotion Classification Models

This study evaluates multiple model variants across all three paradigms: classical machine learning with statistical feature extraction (Naive Bayes, Logistic Regression, LinearSVC, XGBoost, MLP), deep learning with pre-trained word embeddings (Bidirectional GRU and LSTM with GloVe and FastText), and fine-tuned transformer language models (BERT, RoBERTa, DistilBERT). All models are evaluated on a six-class emotion dataset of 20,000 samples under identical experimental conditions. Results show that the transition from statistical features to GloVe-based embeddings produces the largest single performance gain (+3.6 F1 points), while full transformer fine-tuning yields only a marginal additional improvement (+0.8 points) at substantially higher computational cost. The study provides practical model selection guidance for teams working under real-world resource constraints.