An Extensive Experiment on Emotion Classification in Social Media Utilizing Multiple Text Representation Techniques and Deep Recurrent Neural Networks

Emotion classification in text is essential for applications such as sentiment analysis and mental health monitoring, particularly in social media, where users frequently share opinions and emotions. This study evaluates three text representation techniques, namely Term Frequency-Inverse Document Frequency (TF-IDF), Word to Vector (Word2Vec), and Global Vectors for Word Representation (GloVe), combined with four deep recurrent neural networks, including Gated Recurrent Unit (GRU), Long Short-Term Memory (LSTM), Bidirectional GRU (Bi-GRU), and Bidirectional LSTM (Bi-LSTM). Using a dataset of 393,822 samples categorized into six emotions, sadness, joy, love, anger, fear, and surprise, the data underwent preprocessing, including noise removal, spelling correction, and lemmatization. Exploratory Data Analysis (EDA) examined label distribution and linguistic patterns to inform model development. A two-step evaluation process selected models based on validation loss and F1-Score. Word2Vec with Bi-LSTM achieved the highest F1-Score of 0.9248, outperforming other configurations. Pre-trained embeddings like Word2Vec and GloVe demonstrated greater effectiveness than statistical approaches such as TF-IDF when paired with recurrent models.