Araújo, Daniel Sabino Amorim deSilva, Felipe Morais da2022-02-172022-02-172022-02-04SILVA, Felipe Morais da. Análise automatizada de pedidos de recurso a infrações de trânsito utilizando Processamento de Linguagem Natural. 2022. 41 f. Trabalho de Conclusão de Curso (Graduação em Ciência da Computação) – Departamento de Informática e Matemática Aplicada, Centro de Ciências Exatas e da Terra, Universidade Federal do Rio Grande do Norte, Natal, 2022.https://repositorio.ufrn.br/handle/123456789/46060Over the years, the need to automate repetitive processes grew as new market demands emerged and, eventually, new technologies were developed in order to make this demand viable. Computing, in particular, is an area that was responsible for creating countless techniques capable of automating daily tasks and, nowadays, it is gaining even more space thanks to the advances made in recent years by Artificial Intelligence. Furthermore, there is a sub-area of Artificial Intelligence, which is Machine Learning, whose objective is to create models capable of representing a given problem in order to obtain answers to specific questions. Bringing this discussion to a real scenario, the Polícia Rodoviária Federal (PRF) of Rio Grande do Norte, among its various activities, needs to analyze a large amount of resources from fines in order to guarantee the citizen’s right to contest an eventual injustice in law enforcement. These fines can be submitted by completing physical or electronic documentation. Additionally, there is no standardization in the structure of the document, since both common people and legal professionals can produce this document. After the responsible agent analyzes the appeal, it is necessary to justify the approval or rejection regarding the document. In order to assist and expedite the analysis of fines resources, which is a repetitive and tiring process, the objective of this work arose in conjunction with the PRF of RN, which is to develop a system capable of, given a fine resource as input , process your content and provide a pre-opinion so that the reviewer can save a great deal of time writing the appeal justification. To develop this solution, PRF made available more than a thousand documents, which went through a text extraction and data pre-processing process. After that, it was planned to develop a processing pipeline to: (i) transform texts into numerical and/or vector representations; (ii) use numerical and/or vector representations to build intelligent models based on Machine Learning; (iii) evaluate and compare the results obtained from different techniques. The models were trained: Naive-Bayes and SVM using the Bag of Words (BOW) and Term Frequency–Inverse Document Frequency (TF-IDF) representations; Long Short-Term Memory (LSTM) with the Word2Vec representation. After implementing these models, they were evaluated in terms of accuracy, precision, recall and F-score metrics. It was noticed that the models with the best performance in relation to these metrics were, respectively, SVM with TFIDF and LSTM with Word2Vec. Finally, a REST API was developed to query these models.Attribution 3.0 Brazilhttp://creativecommons.org/licenses/by/3.0/br/Recurso de multaProcessamento de Linguagem NaturalAprendizado de MáquinaNatural Language ProcessingMachine LearningComputaçãoAppeal to contest finebachelorThesis