Bridging the Gap Between Traditional Machine Learning and Advanced Deep Learning in Emotion Detection: A Comprehensive Analysis Using Standardized Metrics on Multi-Class Text Classification

Abstract

Emotion detection from textual data is a critical task in natural language processing (NLP), with applications spanning mental health monitoring, customer service, social media analysis, and human-computer interaction. This study conducts a comprehensive comparative analysis of machine learning (ML) and deep learning (DL) approaches for multi-class emotion classification using a recently developed dataset containing seven emotion categories, including the often-overlooked neutral class. We applied rigorous preprocessing-text normalization, tokenization, stopword removal, and vectorization-to ensure consistent input across all models. Seven classical ML algorithms (e.g., Support Vector Machine, Logistic Regression, Gradient Boosting, and Multinomial Naive Bayes) and seven DL models (e.g., LSTM, BiLSTM, GRU, BiGRU, CNN, CNN-BiLSTM, and CNN-Attention) were evaluated under the same conditions using standardized performance metrics (accuracy, precision, recall, and F1-score). The results highlight that while Linear SVM achieved strong performance among ML models (87.3% accuracy), the 1D CNN and BiGRU consistently outperformed others in the DL category (up to 93.5% accuracy). These findings emphasize the importance of model architecture and data preprocessing in capturing nuanced emotional patterns in text. The proposed framework provides a unified benchmark for future research and practical deployments of emotion-aware NLP systems.