Fine-Tuning Arabic and Multilingual BERT Models for Crime Classification to Support Law Enforcement and Crime Prevention

Abstract

Safety and security are essential to social stability since their absence disrupts economic, social, and political structures and weakens basic human needs. A secure environment promotes development, social cohesion, and well-being, making national resilience and advancement crucial. Law enforcement struggles with rising crime, population density, and technology. Time and effort are required to analyze and utilize data. This study employs AI to classify Arabic text to detect criminal activity. Recent transformer methods, such as Bidirectional Encoder Representation Form Transformer (BERT) models, have shown promise in NLP applications, including text classification. Applying these models to crime prevention motivates significant insights. They are effective because of their unique architecture, especially their capacity to handle text in both left and right contexts after pre-training on massive data. The limited number of crime field studies that employ the BERT transformer and the limited availability of Arabic crime datasets are the primary concerns with the previous studies. This study creates its own X (previously Twitter) dataset. Next, the tweets will be pre-processed, data imbalance addressed, and BERT-based models fine-tuned using six Arabic BERT models and three multilingual models to classify criminal tweets and assess optimal variation. Findings demonstrate that Arabic models are more effective than multilingual models. MARBERT, the best Arabic model, surpasses the outcomes of previous studies by achieving an accuracy and F1-score of 93%. However, mBERT is the best multilingual model with an F1-score and accuracy of 89%. This emphasizes the efficacy of MARBERT in the classification of Arabic criminal text and illustrates its potential to assist in the prevention of crime and the defense of national security.