Abstract
With the growth of social platforms in recent years and the rapid increase in the means of communication through these platforms, a significant amount of textual data is available that contains an abundance of individuals’ opinions. Sentiment analysis is a task that supports companies and organizations to evaluate this textual data with the intention of understanding people’s thoughts concerning services or products. Most previous research in Arabic sentiment analysis relies on word frequencies, lexicons, or black box methods to determine the sentiment of a sentence. It should be noted that these approaches do not take into account the semantic relations and dependencies between words. In this work, we propose a framework that incorporates Arabic dependency-based rules and deep learning models. Dependency-based rules are created by using linguistic patterns to map the meaning of words to concepts in the dependency structure of a sentence. By examining the dependent words in a sentence, the general sentiment is revealed. In the first stage of sentiment classification, the dependency grammar rules are used. If the rules are unsuccessful in classifying the sentiment, the algorithm then applies deep neural networks (DNNs). Three DNN models were employed, namely LSTM, BiLSTM, and CNN, and several Arabic benchmark datasets were used for sentiment analysis. The performance results of the proposed framework show a greater improvement in terms of accuracy and F1 score and they outperform the state-of-the-art approaches in Arabic sentiment analysis.
•Innovative dependency rule-based approach for Arabic sentiment analysis. These rules overcome the limitation of word frequency based approaches by employing linguistic patterns that permit the sentiment to transfer from words to concepts based on the dependency structure of a sentence. Furthermore, these rules are fully explainable and explore the terms and dependencies more comprehensively to provide a justification for each production. Thus, understanding the model predictions in an interpretable way can provide trust and transparency.•A comparative analysis of the proposed hybrid Arabic Sentiment Analysis Framework with Logistic Regression, Support Vector Machine, Convolutional Neural Network, Long–Short Term Memory and Bidirectional LSTM.•An ablation study of the proposed dependency rule-based approach on several datasets illustrating the importance of each rule.•Overcoming the limitation of unclassified reviews using Arabic dependency rule-based approach by combining the rules with DNN models.