Crime Text Classification and Drug Modeling from Bengali News Articles : A Transformer Network-Based Deep Learning Approach

In recent years, major cities in Bangladesh have witnessed a concerning rise in drug-related activities. The prevalence of drug-related crimes is evident through daily newspaper reports highlighting incidents stemming from drug use, tracking, and addiction. Consequently, public safety has emerged as a paramount concern for the government [1]. This study proposed a transformer network based deep learning technique for Crime Text Classification and Drug Modeling from Bengali News Articles. We initiated by preprocessing our dataset, followed by feature extraction using the Hugging Face transformer BERT. The derived features were subsequently fed into a binary classifier to discern between crime and non-crime text. Subsequently, crime text undergoes further classification into various crime types. Utilizing a Transformer Network-Based Deep Learning Approach, specifically employing zero-shot classification techniques for both binary and multiclass scenarios, we extracted drug-related crime content from Bengali newspapers. The absence of labeled data posed a formidable challenge, prompting the application of Named Entity Recognition (NER) techniques for the identification of geographical locations within the drug-related news articles. Notably, our results emphasize outstanding performance for both binary classification (crime and non-crime categories) with commendable precision(96%), recall(92%), and F1-scores(94%), as well as multiclass classification (different types of crimes) with elevated precision(95%), recall(98%), and F1-scores(97%) in the Bengali text. This research bears significant implications for crime detection, prevention, public health, and drug enforcement. The findings indicate that transformer network-based deep learning approaches, such as the zero-shot technique, can effectively classify Bengali newspaper articles, showcasing their potential for diverse applications, including drug modeling.