Towards Emotionally Aware Artificial Intelligence: A Comprehensive Analysis of Classical and Deep Learning Approaches for Multiclass Emotion Detection in Text

Abstract

Emotion recognition from user texts has become increasingly vital in understanding human views, improving human-computer interaction, and strengthening psychological and social applications. With the growing volume of usergenerated content online, identifying the underlying emotional states can contribute significantly to areas such as mental health monitoring, marketing, and AI-based conversational agents. However, accurately classifying emotions remains a challenging task due to the complexity of natural language, subtle differences in expression, and imbalanced data sets. This research aims to understand the effectiveness of multiple machine learning algorithms including Support Vector Machines (SVM), Random Forest, Naive Bayes, XGBoost, and deep learning models such as BiLSTM and DistilBERT in emotion classification. The main objective is to efficiently preprocess textual data and compare the performance of ML models and deep learning models. Our results show that RoBERTa achieved the highest performance among deep learning models, while LightGBM outperformed other traditional machine learning methods, demonstrating their effectiveness in accurately classifying emotions in text. From the discussion, it became clear that although classical models are easier and faster to train, deep learning models, especially transformers, are much better at capturing the deeper emotional meaning in text.