Análise de sentimentos para o português usando redes neurais recursivas

Abstract

In this work we present a trained model for sentiment analysis in Brazilian Portuguese sentences. We used a recursive neural tensor network implementation developed by the University of Stanford team that achieved state-of-the-art results on sentiment analysis on English sentences. A sentiment Treebank for sentiment analysis in Portuguese was formed based on a previous data set of book reviews. The Treebank is formed by 12.512 sentences in Penn Treebank format containing associated polarities for each sentence distributed in three classes - Positive, Neutral and Negative. In order to generate the Treebank, syntactic trees were extracted from the corpus sentences using parsers developed by the University of Stanford and the University of Berkeley trained with a Portuguese data set. Our work describes the construction of the corpus and the use of it on the training of a sentiment analysis model. This model was tested following 10-fold cross-validation method and obtained 51, 18% of accuracy on positive and negative prediction and 69, 08% of accuracy on the prediction of all the sentences.