EURO 2024 Copenhagen
Abstract Submission

3503. Dealing with noisy labels in text analytics: an application in crime analytics for prioritizing reports

Invited abstract in session MB-31: Crime Analytics, stream Analytics.

Monday, 10:30-12:00
Room: 046 (building: 208)

Authors (first author is the speaker)

1. Carla Vairetti
Universidad de los Andes
2. Sebastian Maldonado
Department of Management Control and Information Systems, University of Chile
3. Richard Weber
Department of Industrial Engineering, FCFM, University of Chile

Abstract

Deep learning has emerged as the predominant method for text analytics owing to its capacity to learn language patterns from vast datasets and subsequently apply this knowledge to specific tasks. In this study, we leverage deep learning techniques to enhance the categorization of public security reports submitted by ordinary individuals through a mobile application in Chile. Users select a category to report an incident from a wide array of options, encompassing everything from lost pets and disturbing noises to accidents involving injuries or burglary. They also have the option to provide a description of the incident. A primary challenge arises from users often misreporting events in the first category presented in the app.
While the app has proven invaluable for municipalities and police departments in efficiently deploying resources and patrols to address incidents and prevent crime, its prioritization scheme relies on the category provided by the user. This study introduces a machine learning solution employing BERT and other Transformer architectures to learn from incident descriptions and accurately assign labels, thereby improving event prioritization. Our results show a positive predictive performance. Additionally, the model shows potential to extract further insights from descriptions, augmenting prioritization by, for instance, assessing the severity of accidents or crimes.

Keywords

Status: accepted

Back to the list of papers