Fine-grained emotion detection from text: A comparative study of classical ml, lstm, hybrid ensembles, and transformer approaches

Abstract

Emotion detection from text has emerged as a critical task in sentiment analysis, customer analytics, social media monitoring, mental-health assessment, and human–computer interaction. However, identifying fine-grained emotions from short and contextually ambiguous text remains a challenging problem. This study proposes a comprehensive framework that integrates classical machine learning, deep learning, and transformer-based approaches for sentence-level emotion classification. The methodology includes tf–idf–based random forest, an embedding-driven lstm model, a novel hybrid ensemble combining random forest, adaboost, and gradient boosting, and a fine-tuned bert model as a modern contextual baseline. Experiments were conducted on a benchmark kaggle emotion dataset, and performance was evaluated using accuracy, macro-precision, macro-recall, and macro-f1. Results show that the proposed hybrid ensemble achieves the highest performance with 94.6% accuracy, outperforming both the lstm (85.63%) and the fine-tuned bert model (89.8%). The study further provides comparative insights across feature-engineering strategies, contextual embeddings, and ensemble learning. The findings demonstrate that the hybrid ensemble captures discriminative emotional cues more effectively than individual classical or deep learning models, offering a reliable and high-performing solution for real-world text-based emotion detection applications.