Acurácia de um modelo fonotático de entropia máxima aplicado ao português brasileiro

• Ano de defesa: 2017
• Autor(a) principal: Alves, Fernando Cabral
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal da Paraíba; Brasil; Linguística; Programa de Pós-Graduação em Linguística; UFPB
• Palavras-chave em Português: Fonotática; Entropia máxima; Aprendizado de máquina; Estimação de magnitude; Phonotactics; Maximum entropy; Machine learning; Magnitude estimation; CNPQ::LINGUISTICA, LETRAS E ARTES::LINGUISTICA
• Link de acesso: https://repositorio.ufpb.br/jspui/handle/123456789/14271

The present work is part of the studies that seek to represent and investigate linguistic systems using mathematical models. In such context, a Maximum Entropy model of phonotactics developed by Hayes and Wilson (2008) has exhibted a high level of accuracy in relation to experimental data when applied to English, outperforming other phonotactic modelling proposals. Nevertheless, despite its good results, we are ignorant of any work in Brazil which makes use of the model or of Maximum Entropy models in general. Since the model is universal (i.e. applicable to any language), we have taken our objective to be measuring the level of accuracy of the model when applying it to Brazilian Portuguese. The text is divided into three chapters. In the first chapter, we have described in details the model to be tested. In the second one, we have presented the methodology employed to: i) apply the phonotactic model to Brazilian Portuguese; and ii) collect experimental data against which we measure the accuracy of the model predictions obtained in i). The methodological procedures involved the creation of two softwares, one for automated phonological transcription of Brazilian Portuguese and a second one for carrying out magnitude estimation experiments. Finally, in chapter three we show the results. In two applications, the correlation between model predictions and experimental data, measured by the Pearson coefficient, were found to be in the region of 0 and 0,5, thus showing a much weaker linear dependence than that found for English (0,946).