Authors: Bhagyashri Kasar, Ranjana Dahake
Abstract: Conventional predictive policing methods struggle to handle the vast and diverse data gen-erated by urban crime environments. Rule-based and traditional machine-learning models often struggle to identify contextual significance, adapt to evolving crime patterns, and maintain stable performance when faced with data imbalances. The use of large language models makes semantic intelligence a revolutionary concept, as it goes beyond the mere structured attributes and gains the reasoning that is deciphered. The current research uncov-ers a model of smart policing that is aware of language and is able to predict the types of crimes accurately with the help of the semantic understanding acquired from old incident records, spatial-temporal features, and text descriptions. The methodology investigates prompt-driven reasoning strategies like zero-shot, few-shot, and task-adaptive inference to recognize crimes without needing a lot of retraining or manual feature engineering. Com-parative analysis with traditional predictive models provided insights regarding the ad-vancements in adaptability, interpretability, and minority class recognition. The findings indicate that the use of semantics in the form of intelligence has enhanced the prediction of crime types and made support for public security operations more flexible, scalable, and context-sensitive. Linguistic-based crime analytics can significantly assist police agencies in their efforts to anticipate incidents, allocate manpower effectively, and implement data-driven policing strategies in various urban areas.
DOI: http://doi.org/10.5281/zenodo.20848323
