Abstract:
The paper examines the effectiveness of multi-class sentiment analysis strategies using deep learning methods for
imbalanced and balanced datasets with and without word embeddings. Seven models, including Convolutional Neural Network
(CNN), Long Short-Term Memory (LSTM), CNN-LSTM, LSTM-CNN, Bidirectional LSTM (BILSTM), CNN-BILSTM and Two
layers CNN are compared. A dataset consisting of 23,168 tweets was gathered from online learning platforms between 2020 and
2021. The performance of sentiment categorization was evaluated based on accuracy, precision, recall, and F1-score. The study
presents three main findings: (1) a comparison of the effectiveness of seven sentiment analysis algorithms, (2) the clear advantage of
pre-trained Word2vec, and (3) the capability to achieve a balanced sentiment categorization using Twitter data. The LSTM-CNN
model utilizing Word2vec word embedding outperformed several models, achieving an accuracy of 89.66% and a precision, recall,
and F1-Score of 90.00% for the testing results. The experimental results confirmed that this methodology enhanced the accuracy of
sentiment classification compared to standard methods and exhibited superior classification performance. The empirical research
showed that the LSTM-CNN method was fast, efficient, and viable, making it a potentially better option for optimizing online
learning rules. The study provides valuable insights to analytics professionals and academicians engaged in text analysis. It focuses
on the performance evaluation of essential algorithms in sentiment classification, particularly emphasizing the data balancing
technique in deep learning hybrid models.
