Classificação de documentos do exército brasileiro utilizando o classificador Naive Bayes e técnicas de seleção de sentenças

Abstract

One of the needs of the Brazilian Army is to perform the classification of documents called “Boletins Internos”, which must be grouped in order to generate summarized reports about the military. To accomplish this, it is necessary to find relevant references to each military inside a set of documents, elaborated during the period of one semester. To perform this classification automatically, this work uses the Bayes classifier. The classifier employs n-grams as a way to select the training attributes, identifying the frequency/occurrence of words inside the analyzed documents. It is also necessary to identify which sentences of the documents are related to the analyzed military. We propose two heuristics in order to better perform the selection of sentences that are related to each military. We can see that the proposed implementation can achieve 78.5% F-Measure in the recovery of relevant documents. Furthermore, the use of n-grams can perform a more accurate analysis of the information, and the sentence selection directly influences the classification.