Aprimoramento da análise de sentimentos em redes sociais utilizando análise léxica e perfil de usuário

Abstract

Social networks have a large amount of data available to be exploited in an increasingly comprehensive way. Sentiment analysis uses this data to develop useful applications in people’s daily lives. However, there are still failures in this type of analysis, either by the restricted number of words contained in dictionaries or by not considering all parameters that can influence the final sentiment of a sentence such as: users profile, punctuation, geographical location or social networks using frequency, for example. More reliable results can be obtained by taking advantage of the greater number of these parameters or by grouping the most suitable parameters. First, this work suggests that the result of sentiment analysis is influenced by the adverbs punctuation, whose proposal is to reverse or intensify the final sentiment of a sentence depending of the adverbs. We work on the proposal to consider user profile characteristics such as age and gender to determine the sentiment value of each sentence posted on a social network. A Recommendation System (SR) was presented, based on the sentiment analysis of sentences extracted from social networks, from an algorithm that considers the adverbs punctuation. We also performed a detailed analysis with 7000 sentences to determine which characteristics would be more relevant, such as punctuation, number of characters, media sharing, subjects, among others; and which characteristic could be disregarded. Different machine learning algorithms were tested in search of the best result for classifying the users by age group. Through the punctuation of the sentences considering the adverbs, it was possible to obtain the absolute maximum error corresponding to 0.21, an inferior value compared to the results presented by other tools of sentiment analysis. In order to classify users by age group, the Deep Convolutional Neural Network (DCNN) had the best performance, reaching an accuracy of 0.95 in the validation tests. In addition, to validate the utility of the proposed model to classify age groups, the model was implemented in the Enhanced Sentimeter Metric (eSM), and the eSM metric results when the age group information was not available, were improved with the proposal of age group classification of this work. The advances and tests carried out show that the tools for sentiment analysis in texts extracted from social networks presented increasingly reliable and realistic results.