Comparison Performance of Long Short-Term Memory and Convolution Neural Network Variants on Online Learning Tweet Sentiment Analysis

Abstract

Sentiment analysis can be act as an assisted tool in improving the quality of online teaching and learning between teachers and students. Twitter social media platform currently more than 500 million tweets sent each day which is equal to 5787 tweets per second. Therefore, it is hard to track users’ overall opinions on the topics contained in social media. To catch up with the feedback on online learning, it is crucial to detect the topic being discussed and classify users’ sentiments towards those topics. Even though there are many approaches in developing sentiment analysis models, DL models prove to provide the best performance in the sentiment analysis field. Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) are two mainstream models in DL used for sentiment analysis classification. Therefore, we evaluate CNN, LSTM, and its hybrids to classify sentiment or an online learning tweet from 2020 until 2021 of 23168 tweets. CNN-LSTM, LSTM-CNN, Bidirectional LSTM, CNN-Bidirectional LSTM models were designed and evaluated based on random hyperparameter tuning. We explain the proposed methodology and model design illustration. The outcome assesses the superiority of all models with a remarkable improvement of accuracy and a reduction loss when applying the random oversampling technique. Specifically, the LSTM-CNN model with random oversampling technique outperformed the other six models with an accuracy of 87.40% and loss value of 0.3432. However, the computational time has resulted increased when with random oversampling technique. Thus, in the future, the performance can be improved on computational time and hyperparameter selection with the employment of nature-inspired computing for fast and optimal results.