Aprendizado profundo para reconhecimento de entidades nomeadas em narrativas de roubos

Abstract

Different forms of violence have shown a marked growth spiral worldwide. Since the first decade of the 21st century, the Brazilian federal government and some state governments have been promoting new policies to combat violent crime. To monitor the evolution of crimes and violence, as well as promote the population’s access to public security information in the state of Ceará (Brazil), the Secretariat of Public Security and Social Defense presents its statistics monthly. The data used to collect these statistics have different performance indicators for criminal analysis. One of the key-indicators, called Violent Crimes Against Patrimony (CVP in the Portuguese acronym), encompasses all crimes classified as theft, except robbery-homicide. The occurrence rates for this indicator are high and have not been decreasing over time. Finding and disseminating relevant and timely information in CVP Police reports is crucial for the population and play a central role in the capabilities to combat types of theft. However, analyzing this large volume of data requires manual and extensive work. One way to automate this process is to use the Named Entity Recognition (NER) task for information extraction. In recent years, Deep Learning, empowered by continuous real-valued vector representations, has been widely employed in NER systems. This work proposes the creation of a Deep Learning based NER model capable of recognizing named entities in CVP Police reports, following three main research questions. The research questions define a broad experimental protocol and explore problematic aspects present in the domain, such as data imbalance, number of named entities, and embedding representation (CVP2Vec). The results show high performance, with the Balanced Accuracy metric reaching 88.1% and the F1-score with 82.9% for the best models, outperforming the baseline.