Compilação, anotação e análise linguístico-computacional de um corpus de textos literários dos séculos XIX e XX: corpus Coelho Neto

Abstract

This thesis is the compilation, morphosyntactic annotation and linguistic and computational analysis of a corpus of literary texts of 19th and 20th centuries: Corpus Coelho Netto (CCN), containing texts of the novels A Conquista and Turbilhão and short stories of the book Sertão. The work is in the Corpus Linguistics and Computational Linguistics interface (BERBER SARDINHA, 2000, 2003, 2004, 2005, 2009; BERBER SARDINHA; ALMEIDA, 2008; OLIVEIRA, 2009; BIDERMAN, 1998, 2001; ALUÍSIO; ALMEIDA, 2006; SHEPHERD, 2012; MACENERY AND WILSON, 2001; LEECH, 2004; ALVES; TAGNIN, 2012; ALENCAR, 2009, 2010a, 2010b, 2011a, 2011b, 2013a, 2013b). The CCN contains 53.080 (fifty-three thousand and eighty) tokens. The compilation consists of the steps selection, collection off texts and handling; in which cleaning, editing and updating of texts (ALUÍSIO; ALMEIDA, 2006), and then be submitted to the morphosyntactic annotation and linguistic-computational analysis, with the goal of obtaining data to show whether or not the "excessive" use of adjectives, verbs and adverbs in “–mente”, demonstrating the lexical diversity in Coelho Netto´s texts, noting if what the modernist critics said about the writer was correct. The annotation was performed by automatic tagger Aelius, AeliusHunPos model, free software in Python that uses the Natural Language Toolkit – NLTK library (BIRD; KLEIN; LOPER, 2009), in the pre-processing of texts, in the construction of morphosyntactic tagger and the automatic annotation of corpora with the help of human review (ALENCAR, 2010a, 2013a, 2013b), and it was trained in the Historical Corpus of Tycho Brahe Portuguese (CHPTB). The compilation and annotation CCN involves other actions such as revaluation the accuracy of this tagger in literary texts. The search results indicated that: AeliusHunpos demonstrated better performance than other texts already noted (97.9 %); AeliusHunPos model showed a far beyond performance by annotating corpora with AeliusMaxEnt model; and that, after selection and manual correction of 10% annotated corpora and generated gold standard files, it is suggested an improvement of the approximate 3% of errors by the tagger, in order to increase its accuracy. Regarding the analyzes performed with the CCN, it was found that: lexical diversity - about verbs, adjectives and adverbs in “–mente” considered exaggerated by critics to Coelho Netto unfounded, because his texts are rich, but when compared to the texts by Aluísio Azevedo and Camilo Castelo Branco, comparison of corpus, present vocabulary richness similar to CCN, as exposed in the results.