A Code-Switched Arabic-English Sentiment Analysis Approach Based on Deep-Learning
Tasneem S. Almasah, Gamal A. Ebrahim, Marwa A. Abdelaal · 2023
Sentiment analysis within Online Social Networks (OSNs) becomes a major challenge. Mainly, because of the large amount of data on social networks and the mix of different languages that can be used in these environments. Due to the lack of existing research and datasets on code-switched Arabic-English sentiment analysis, this study creates a new dataset collected from various online social networks. The data is subjected to several preprocessing steps before being analyzed using various deep learning architectures. Moreover, a thorough evaluation of the preprocessing techniques and the performance of the deep learning models have been conducted. This is achieved using comprehensive Twitter datasets in Arabic and English. The study reveals that individual deep learning architectures, specifically, Convolutional Neural Networks (CNNs), Bidirectional Long Short Term Memory (Bi-LSTM), and LSTM demonstrated superior results compared to their hybrid counterparts. Notably, CNN emerged as the most accurate model with an accuracy of 83%, followed closely by Bi-LSTM at 82.3%. Additionally, the paper presents the deployment of an automated machine learning tool, which is Auto-Keras. Furthermore, it designs an optimized sentiment analysis model that aims to enhance accuracy. This approach leads to an increase in the model accuracy to 86%.