Método para analisar autoria de textos baseado em regras de associação e redes de palavras

Abstract

This thesis deals with methods for authorship analysis and verification. More specifi cally, authorship analysis (AA) of digital texts, in particular, literary works in Portuguese from Brazil and Portugal. The AA investigates how similar is a work by an unknown author to a set of works by a known author (LAGUTINA et al., 2019; ROCHA et al., 2017; VENCKAUSKAS et al., 2015; BOUANANI; KASSOU, 2014; TAMBOLI; PRA SAD, 2013; STAMATATOS, 2009; HOLMES, 1985). Authenticating the real author of a work is important, it prevents false authorship attributions and prevents, for exam ple, that the notoriety of a writer is used to spread ideas that, originally, are not his. To analyze the similarities between the works, the AA extracts, organizes and compares structures that appear in the texts, such as number of letters, length of sentences, re petition of words, etc. (ROCHA, 2019; JAMIL; MUSTAFA, 2018; REXHA et al., 2007; MARKOV; BAPTISTA; PICHARDO-LAGUNAS, 2017; VOROBEVA, 2016). In AA, we identified a lack of studies that verify the authorship of a text from combinations of words (i.e. sets of words that recurrently appear among the sentences of a writer and that do not appear in the sets of words of other writers). In this sense, this thesis aims to present a new method for authorship verification. Word combinations are assumed not to happen randomly; they occur in accordance with the syntactic and semantic knowledge that the authors have and evidence of their language (CHOMSKY, 2018; CHOMSKY, 1994; MI OTO; SILVA; LOPES, 2007; FRANCHI; NEGRAO; MULLER, 1998). The advantage of analyzing word combinations lies in discovering patterns related to each author and the contexts of production and publication of each work. The method proposed in this thesis extracts word combinations through association rules, consolidates the combinations into word networks and, from sixteen network metrics, analyzes and infers, in literary works, the periods of editions, the varieties of language Portuguese used, the literary schools and the authors. In this sense, this thesis contributes to AA with a working method that, in addition to verifying authorship, highlights the contexts that, supposedly, an unpublished text by an author should present.