Utilizando um modelo transformer no processo de identificação de entidades nomeadas em textos criminais

Abstract

Identifying essential information makes it easier to analyze documents such as criminal records. The Secretariat of Public Security and Social Defense of Ceará publishes monthly its statistics separated by criminal indicators to monitor the progress of crime and violence. One of the main indicators is the Lethal and Intentional Violent Crimes (CVLI in the Portuguese acronym), which have daily reports released with important information. The extraction of relevant information uses a Natural Language Processing technique called Named Entity Recognition (NER), which involves locating and classifying named entities in texts. NER models are usually based on convolutional or recurrent neural networks. However, Transformers are becoming a new paradigm of NER. The Transformer model facilitates parallelization during training, enabling training on larger data sets than was possible before it was introduced. From this, pre-trained systems were developed, such as BERT (Bidirectional Encoder Representations from Transformers), trained with a massive set of general language data, and can be tuned for language-specific tasks, such as NER. This work proposes the production of a NER model using stratified data, capable of recognizing named entities in CVLI texts, with the Transformer model pre-trained in Portuguese. The model obtained good results. The metric F1-score obtained the value of 71.44% and the Balanced Accuracy of 66.73%, surpassing the baselines.