Developing Arabic Sentiment Analysis for Saudi Arabia's Telecommunication Companies using Deep and Ensemble Learning

Resumen

Sentiment analysis is a type of artificial intelligence that uses algorithms to determine whether an opinion is positive or negative. Arabic Sentiment Analysis (ASA) is responsible for assessing people’s opinions, feelings, and responses to a variety of products and services on social and commercial networking sites. In this article, we develop a new Arabic sentiment analysis model for Saudi Arabia's telecommunication companies (Zain, Mobily, and STC). We create and develope a new dataset, called  Sara-Dataset, for analyzing customer opinions towards Saudi Arabian communication firms. using Google Colabs libraries (e.g., Keras), our dataset consists of 50532 tweets. After processing and preparation it becomes 27294 tweets. The dataset is divided into three parts: training, validation, and testing, which each represent 70\%, 10\%, and 20\%, respectively. The proposed model depends on deep learning models: LSTM (long-short-term memory) and CNN (conventional neural network). We evaluate our model using several parameters. The number of training epochs, the loss function, the optimizer (Adadelta, Adagrad, and Adam), batch size, and the ensemble learning approach We evaluate our model in terms of accuracy, precision, F-measure, R-call, ROC, and loss function. When compared to other models, our results indicate significant enhancements. The best accuracy results of the LSTM model with the Adam optimizer and 32-batch size are 94.97\%. The best accuracy result of the CNN model with the Adam optimizer and 32-batch size is 96.83\%. Our future work is to use ensemble learning model.