Abstract

This study tackles the challenge of noisy labels in crime analytics by developing a Transformer-based deep learning model that corrects misclassified user reports from a Chilean crime app. We selected this architecture for its advanced capabilities in understanding semantic nuance and context, which is essential for handling the noisy, unstructured text of citizen reports. Our method, validated on a real-world dataset of over 1.5 million user reports from 2019-2021, employs an iterative, semi-supervised framework (Algorithm 1) that first identifies a small, clean set of reports using expert-defined keywords, then progressively corrects and incorporates high-confidence predictions from the larger noisy dataset. Additionally, we introduce an automatic prioritization scheme-a hybrid expert system that combines our model's corrected labels with a rule-based keyword matcher-that utilizes this label-correcting model to enhance real-time crime response. The outcome of this research is an AI-powered expert system designed to optimize police resource allocation and municipal crime prevention strategies. To the best of our knowledge, this is the first study to apply deep learning to crime-tracking app data with a focus on label correction and response optimization. The performance of this multiclass label correction model is evaluated using F1-micro and F1-macro scores. Our results show the effectiveness of the proposed system, achieving 95.68% accuracy in classifying urgent and non-urgent user reports, as detailed in a full confusion matrix.