Aplicação de algoritmos na análise de estruturas com verbos psicológicos

Abstract

This dissertation investigates the syntax-semantics interface by analyzing the mapping of thematic roles in structures involving psychological verbs in Brazilian Portuguese. The central objective is to evaluate the performance of machine learning algorithms in the automatic identification of these roles, given the inherent complexity of these predicates which, by expressing mental states, present a variable argument configuration between Experiencer and Theme. Theoretically, the study is anchored in the Generative tradition and the taxonomy proposed by Cançado (1996), which provides the necessary basis for the categorization of the analyzed structures. The methodology comprised the construction of a stratified corpus, containing 4,000 structures with psychological verbs divided according to Cançado’s (1996) taxonomy, and the use of Natural Language Processing (NLP) techniques via the Stanza library (Peng, 2020) for syntactic dependency extraction. For the classification phase, the Naive Bayes (Vaithyanathan, 2002), Decision Tree (Magerman, 1995), and Support Vector Machines (SVM) (Cunningham, 2006) algorithms were implemented and compared. The results indicate that models based on Decision Trees and SVM demonstrate greater effectiveness in capturing the structural nuances of psychological verbs, revealing that the integration between theoretical linguistic foundations and computational models is essential for the advancement of automatic semantic labeling tasks.