Estudo comparativo de análise de sentimentos aplicado à notícias públicas

• Ano de defesa: 2018
• Autor(a) principal: CARVALHO, Caio Magno Aguiar de
• Orientador(a): BARROS FILHO, Allan Kardec Duailibe
• Banca de defesa: FONSECA, João Viana , AGUILAR, Paulo Armando Cavalcante
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA DE ELETRICIDADE/CCET
• Departamento: DEPARTAMENTO DE ENGENHARIA DA ELETRICIDADE/CCET
• País: Brasil
• Palavras-chave em Português: Processamento de Linguagem Natural; Análise de Sentimento; Aprendizagem de Máquina; Seleção de atributos
• Área do conhecimento CNPq: Sistemas de Informação
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/tede/2110

In the elections period, the public opinion about parties and candidates is partially influenced by mainstream media as TV, radio, newspappers and mainly internet through newswire media. However, there is a debate about the impartiality in these media when transmitting news. Sometimes it is acused to favour some political entities and its agendas, while others affirm its neutrality. Assess news article in this context is not a simple task, because the evaluation could be influenced by some biases of who assesses that article. The methods provided by Natural Language Processing, through the field of Sentiment Analysis, could bring a less biased viewpoint of that question. Sentiment Analysis joins text mining techiniques and machine learning tools to classify texts according its sentiment polarity (positive, negative or neutral).In this work we propose a comparative study between sentiment analysis text representation models, feature selection techiniques and machine learning classifiers in order to classify the polarity of political online news about 2014 brazilian elections. In this study the classifiers Naïve Bayes, Support Vector Machine and Logistic Regression (MaxEnt) are evaluated with feature selection techinques as Chi Square, Categorical Proportional Difference, Categorical Probability Propotional Difference. The experiments sought to choose the best text representation, feature selection techinique and machine learning classifier. The evaluation is made by cross validation measuring accuracy mean and its standard deviation. The experimental results pointed to the bag-of-words representation with unibigram selected by Categorical Probability Proportional Difference with MaxEnt classifier achieving 84.45% with standard deviation of 0.029.