Modelo de análise de sentimentos com base na estrutura linguística da sentença

Abstract

Sentiment Analysis (SA) is an area of Natural Language Processing that leads to detect the presence of positive, negative or neutral polarity in a text. SA became an element of greater interest due to the increase of data generation from the internet, and the efforts required to process such volume of data. Due to this demand, there is an endeavour to achieve precise results that will allow the sentiment inserted in the text to be automatically interpreted . Several methods can be applied in this task, however, in this thesis the aspects and challenges in exploring the linguistic structure of the sentence stand out, for which the solutions are directed to rules with the properties of the grammar that describe the natural language. It is worth noting that studies of solutions based on rules are also motivated by excluding costs that involve data training. A review of the state of the art, shows that there are advances in studies of formal linguistics that contribute to computational linguistics, as rules of a syntactic description of Brazilian Portuguese already available in literature, and have not been explored for sentiment analysis. In particular, works with restricted models of rules and the use of parsers were detected to define the parts of speech, sentence structure, and dependency relations. However, no solutions were found involving a grammar constructed to describe a natural language incorporated into a parser that analyzes the sentence structure specifically characterized by sentiment. The main result obtained from this thesis was a new model of sentiment analysis based on an expandable regular grammar defined by rules of semantic composition. Thus, a parser was developed that identifies sentence structures characterized or not with positive or negative polarity. To meet this solution, a set of relations of sentiment between grammatical categories were studied and developed based on the formal description of the sentence structure. A prototype was implemented to test the application of the model in corpora sentences and later for a comparative evaluation of the results, which presented rates at the same levels obtained by other methods.