An Advanced Hybrid Deep Learning Approach for Large-Scale Emotion Detection in Social Media Text Streams

Abstract

Emotion classification from social media sites like X (previously Twitter) is essential for comprehending human behavior, public mood patterns, and market dynamics. The Twitter Emotion Dataset, with more than 500,000 annotated tweets, serves as a valuable yet complex data source owing to its unstructured and noisy characteristics. To tackle these issues, sophisticated deep learning architectures integrating convolutional, recurrent, and attention methods have been utilized to capture both local and sequential dependencies in text. Various models were assessed, including BERT, RoBERTa, DistilBERT, attention-driven LSTM-CNN, attention-driven CNN-GRU, hybrid attention-driven CNN-LSTM-GRU architectures, and BERT combined with LSTM and GRU layers. Optimization techniques including Adam, RMSprop, Nadam, Adagrad, and SGD were examined. The experimental findings indicate that the BERT + LSTM + GRU model attains superior performance, exhibiting 95.3% accuracy, 95.6% precision, 95.3% recall, and 95.3% F1-score. Explainable AI methodologies, particularly LIME, offer interpretability by emphasizing significant aspects, hence showcasing the model’s efficacy in high-dimensional emotion recognition inside extensive social media datasets.