Spam and Sentiment Detection in Arabic Tweets Using MARBERT Model

Social media platforms like Twitter have become essential channels for customers to voice their satisfaction or complaints regarding services. For large organizations like the Saudi Telecom Company (STC), the sheer volume of incoming tweets makes manual sentiment analysis impossible. This paper addresses the need for an automated, accurate system to process Arabic-language feedback, which is notoriously difficult for machines to interpret due to complex morphology and the prevalence of diverse regional dialects.

The researchers utilized MARBERT, a transformer-based model pre-trained on a massive corpus of Arabic tweets. Unlike standard multilingual BERT models, MARBERT is specifically optimized for the linguistic nuances and informal nature of Arabic social media text. The authors trained and evaluated this model using a dataset of 24,513 tweets, categorized into five distinct classes: positive, negative, neutral, sarcasm, and indeterminate. The study focused on fine-tuning this model to accurately capture the sentiment of STC customers, thereby enabling the company to respond more effectively to user needs.

The study confirms that deep learning models, particularly those based on the BERT architecture, significantly outperform traditional machine learning and lexicon-based approaches for Arabic sentiment analysis. By successfully classifying sarcasm and nuanced sentiment, the proposed MARBERT-based scheme provides a scalable solution for companies to monitor public perception in real-time. This work highlights the importance of using language-specific pre-trained models to bridge the performance gap in Arabic Natural Language Processing (NLP) tasks.