Classificação de tendências políticas em notícias via mineração de texto e redes neurais sem peso

Abstract

Biased news can influence the reader's opinion in subtle ways. Nowadays, due to the unprecedented amount of information created and made available through social media, the identification of biases is increasingly challenging. In the domain of politics, addressing the challenge is particularly relevant. In this dissertation, we seek the identification of polarity in Portuguese political news through the process of textual data mining using the WiSARD Weightless NeuralNetwork, and one of its extensions, the Clus WiSARD. The WiSARD classifier works through a structure of discriminators, where each discriminator is responsible for identifying a class. We assessed polarity using the body and the headline of news published in widely known media vehicles. The obtained results are encouraging, indicating the feasibility of automatic and eficient bias detection. We obtained accuracy of about 90% when using full body news and accuracy of the 75% when considering only headlines. In addition, we also perform a temporal analysis on the political dynamics of bias.