Análise de grafos aplicada a produções textuais de alunos do ensino fundamental e seu potencial preditivo da dislexia do desenvolvimento

• Ano de defesa: 2018
• Autor(a) principal: Luz, Joana Paim da
• Orientador(a): Buchweitz, Augusto
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Pontifícia Universidade Católica do Rio Grande do Sul
• Programa de Pós-Graduação: Programa de Pós-Graduação em Letras
• Departamento: Escola de Humanidades
• País: Brasil
• Palavras-chave em Português: Dislexia; Teoria dos Grafos; Habilidades de escrita; Aprendizado de Máquina
• Área do conhecimento CNPq: LINGUISTICA, LETRAS E ARTES::LETRAS
• Link de acesso: http://tede2.pucrs.br/tede2/handle/tede/8158

Dyslexia is a learning disorder of neurobiological origin, which is characterized by an unexpected difficulty when decoding written texts, due to an unsatisfactory learning of the alphabetical principle and a deficient graphological-phonemic association. The ways dyslexia affects writing skills have been scientifically explored so as to investigate linguistic aspects associated with spelling and classes of words used by subjects who suffer from dyslexia. Quantitative studies focused on measuring the structure of their texts are unknown. The main objective of this study was to identify patterns of textual connectivity in good readers, bad readers and dyslexic children, based on the analysis of graph measures extracted from their texts and Machine Learning techniques. Essentially, it sought to investigate (a) whether the type of transcription of the texts - corrected or original - and the normalization of the graph attributes by the number of words of each text interfere significantly in sorting the children in their fluency and schooling groups; (b) whether there is significant differences among good readers’, bad readers’ and the dyslexic participants’ graph attributes; (c) whether measures of each year of data collection converge to similar values and if they are significantly different among the years; (d) whether the graph attributes obtained with Speech Graphs, when associated with Machine Learning techniques, can predict reading fluency levels and, specifically, developmental dyslexia. The hypotheses for the questions listed were all affirmative. To verify them, texts produced by 181 children and adolescents from the ACERTA Project were transcribed and divided in two experimental groups: Ambulatory (N = 52, all dyslexic) and Schools (N = 129, subdivided into good, medium and bad readers). These transcribed texts served as input for Speech Graphs software, which extracted graph attributes representative of the structure of each text. Descriptive and inferential statistical analyzes revealed (a) the prevalence of significance among graphs attributes extracted from non-normalized original transcripts (63.07% significance between analyzes); (b) patterns of textual connectivity by each reading fluency group and (c) patterns of textual connectivity by year of data collection based on significant differences found in five graphs attributes - nodes, edges, largest connected component, density and average smallest path; (d) 2016 as the best year to sort the children in their reading fluency groups by making use of SVM classifiers, considering they reached the highest accuracy (85%), recall (83%), precision (85%) and F1 score (83%) when sorting good readers and dyslexic children texts, based on 2016 graph measures. These findings provide evidences that indicate the possibility to explore and improve a new methodological frame to assess reading fluency through written texts and based on Graph Theory.