Assessing BERT-based models for Arabic and low-resource languages in crime text classification

The bidirectional encoder representations from Transformers (BERT) has recently attracted considerable attention from researchers and practitioners, demonstrating notable effectiveness in various natural language processing (NLP) tasks, including text classification. This efficacy can be attributed to its unique architectural features, particularly its ability to process text using both left and right context, having been pre-trained on extensive datasets. In the context of the criminal domain, the classification of data is a crucial activity, and Transformers are increasingly recognized for their potential to support law enforcement efforts. BERT has been released in English and Chinese, as well as a multilingual version that accommodates over 100 languages. However, there is a pressing need to analyze the availability and performance of BERT in Arabic and other low-resource languages. This study primarily focuses on analyzing BERT-based models tailored for the Arabic language; however, due to the limited number of existing studies in this area, the research extends to include other low-resource languages. The study evaluates these models' performance in comparison to machine learning (ML), deep learning (DL), and other Transformer models. Furthermore, it assesses the availability of relevant data and examines the effectiveness of BERT-based models in low-resource linguistic contexts. The study concludes with recommendations for future research directions, supported by empirical statistical evidence.