Abstract:
Detecting emotions in textual data is a vital task in natural language processing, with
applications
sentiment
analysis,
recommendation systems, and
humancomputer interaction. Emotion classification plays a crucial role in providing
valuable insights into user sentiments and behaviors. This research investigates the
efficacy of various deep learning models combined with different feature extraction
techniques for emotion classification in text data, using a dataset of news articles. While
past research has primarily focused on either a single feature extraction method or a
limited set of models, this work explores the efficacy of three different feature
extraction techniques—TF-IDF, Word2Vec, and GloVe—integrated with five distinct
deep learning models: CNN, RNN, LSTM, MLP, and Transformer. Unlike previous
studies that often overlook the issue of data imbalance, this research includes specific
steps to balance the dataset, thus enhancing model training and performance.
Furthermore, comprehensive hyperparameter tuning was conducted, adjusting
parameters such as learning rate, batch size, epochs, dropout rate, and dense layer units.
This extensive tuning resulted in a significant accuracy of 79% with the Transformer
model using GloVe embeddings. The combination of exploring multiple feature
extraction methods, addressing data imbalance, and fine-tuning model parameters
distinguishes this study from existing works, providing a more robust benchmark for
future research in emotion classification in natural language processing.
