Analyzing Customer Reviews on Banks Mobile Applications: A Bilingual Approach Using Natural Language Processing for Topic Detection and Sentiment Analysis رسالة ماجستير

Abstract

Mobile banking applications have become a central channel for financial interaction, service delivery, and customer engagement in the Arab world due to the rapid expansion of digital banking services. This study aims to analyze Arabic user sentiment toward mobile banking applications and to identify the main topics that influence user satisfaction and dissatisfaction, in order to support service improvement and decision-making in the banking sector. The study was conducted in 2025 using user reviews collected from Google Play for mobile banking applications affiliated with the top 100 Arab banks listed by The Banker (2024). Data were collected using a Python-based data extraction tool, resulting in more than 361,000 reviews. The dataset was filtered to include Arabic-language reviews only, while preserving the natural country distribution across the MENA region, reflecting real usage patterns and dialectal diversity. This dataset represents the study population, from which a labeled subset was used as the analytical sample. The methodology involved developing a domain-specific Arabic preprocessing pipeline that includes text normalization, dialect handling, negation preservation, and meaningful bigram construction. Several annotation approaches were examined and compared, including manual labeling, rating-based sentiment inference, and GPT-assisted classification. Based on reliability and consistency, manual labeling was adopted as the final reference annotation. Class imbalance was addressed using class weights, SMOTE, and random oversampling. Sentiment classification was evaluated using traditional machine-learning models, an LSTM model, and transformer-based Arabic language models (AraBERT and MARBERT). Topic identification was conducted using a dual strategy combining probabilistic Latent Dirichlet Allocation (LDA) and GPT-based topic generation. The results show that transformer-based Arabic models achieve the highest performance, with MARBERT outperforming all baselines due to its strong ability to capture dialectal and contextual information. Topic analysis indicates that positive reviews mainly emphasize ease of use, convenience, and successful transactions, whereas negative reviews are dominated by reliability related issues such as login failures, unstable updates, slow performance, and application crashes. The study recommends prioritizing application reliability, strengthening pre-release testing, and adopting continuous sentiment monitoring to improve mobile banking services.