Uma abordagem sistêmico-funcional da Análise de Sentimentos em português brasileiro orientada para aplicações multilíngues

Abstract

This thesis introduces a systemic-functional approach (HALLIDAY; MATTHIESSEN, 2014) to Sentiment Analysis in Brazilian Portuguese. It focuses on grammatical patterns that can be associated with characterization of emotions in language. The text sample herein analyzed was retrieved from a corpus of sub headlines of newspaper articles written in Brazilian Portuguese, previously compiled by researchers at PURC/PR and annotated with categories naming emotions as defined by Ekman (1970), namely Sadness, Disgust, Fear, Anger, Neutral, Happiness, and Surprise, by language lecturers at PUC/PR and Federal University of Technology Paraná (UTFPR). The sample was segmented into clauses, which were manually annotated according to grammatical categories pertaining to the three language Metafunctions Interpersonal, Ideational, and Textual as established by Halliday e Matthiessen (2014). Two scripts were developed within the R software and environment (R CORE TEAM, 2018), the first one designed for identification of grammatical patterns in annotated clauses and the second one for cluster analysis of categories naming emotions based on selected categories for each group of annotated clauses. The first script yielded results showing grammatical systems potentially useful to characterize categories naming emotions. The Transitivity system was one of the annotated systems that contributed the most to characterization of categories naming emotions. The system of Deixis and presence or absence of Complements and/or Adjuncts had impact on characterization of categories naming emotions as well. Ideational Theme and Textual Theme pointed out results that also supported characterization of categories naming emotions. There were systems that did not show potential to characterize these categories naming emotions, as the systems of Subject Number and Subject Presumption pertaining to the Interpersonal Metafunction and the system of Interpersonal Theme regarding the Textual Metafunction, as these systems presented predominant choices to most of the annotated clauses. Results yielded by the second script showed that configurations obtained for categories corresponding to emotions of Happiness and Surprise were grammatically similar. Categories corresponding to emotions of Sadness and Disgust had different configurations in comparison to categories that corresponded to other emotions, because they composed an isolated group. Results obtained for category corresponding to emotion of Fear, for category corresponding to absence of emotion, Neutral, and, finally, for category corresponding to emotion of Anger were grammatically similar to those obtained for categories corresponding to emotions that formed the first group Happiness and Surprise. The impact of this study has to do with potential multilingual applications of the grammatical approach herein proposed, which may enhance research on Sentiment Analysis, since grammatical patterns can be useful to support the development of an algorithm for machine learning.