Arabic App Reviews: Analysis and Classification

User opinions and feedback on mobile applications are crucial for application developers, offering insights into issues like bugs, popular features, and enhancement requests. Given the vast number of feedback for each app, it is impractical for developers to manually extract valuable information. To better understand and analyze user opinions, developers can benefit from automatic sentiment analysis and classification of app reviews. Existing research has primarily focused on reviews written in English, with some studies addressing sentiment analysis of Arabic reviews but overlooking the classification task. Given the widespread use and complexity of Arabic compared to English, our work investigates both sentiment analysis and classification of Arabic app reviews. We introduce the AURA ( A pp U ser R eview in A rabic) dataset. AURA dataset has two versions: AURA-Sentiment with 29,700 labeled reviews for sentiment analysis, and AURA-Classification with 2,900 labeled reviews for classification. Using these datasets, we applied deep learning (DL) and natural language processing (NLP) techniques for analyzing and classifying Arabic app reviews. Leveraging the MarBert model, we achieved an F1-score of 0.89 for sentiment analysis and an F1-score of 0.62 for a four-class classification problem. Our findings provide valuable insights and suggest directions for future research.