Redes da textualidade: ciência das redes aplicada à modelagem da coesão lexical

Abstract

This doctoral thesis aimed to develop a generalizable model of lexical cohesion based on network science to evaluate patterns in texts of different genres. Its specific objectives were: (a) to define network-based indices of lexical cohesion to differentiate texts from random sets of sentences, (b) to characterize the topology of the networks to verify emergent patterns and establish analysis criteria, (c) to construct a network-based method for analyzing lexical cohesion, and (d) to analyze the lexical cohesion of different text genres to classify them. Adopting critical rationalism and Grounded Theory, four experiments were conducted based on data mining and network analysis. Experiment 1 defined six indices of lexical cohesion and evaluated them in 60 texts from 6 genres and 60 pseudo-texts. The indices were calculated for each sentence and average indices were calculated for each text and each pseudo-text. The values of the indices were. Three indices differentiated texts from pseudo-texts, and one differentiated text genres. The first 10 to 60 sentences proved sufficient for the analysis of textual cohesion with the proposed method. Experiment 2 investigated the impact of different definitions of elements inserted in the networks’ construction. The definitions did not affect the identification of small-world behavior, calculated by efficiency, but influenced the identification of scale invariance. Experiment 3 defined levels of pseudo-texts and non-texts, comparing their cohesion indices with texts. The need for manual text cleaning was also evaluated, and the edge definition method was revised. The revised method and the use of uncleaned texts proved efficient. Non-texts generated by random tokens and uniform random selection resulted in the most appropriate null model. Experiment 4 replicated the results of Experiment 1 with 870 texts and 145 non-texts using the Method for Lexical Cohesion Analysis Based on Network Science initiated in Experiment 1 and refined in Experiments 2 and 3. The average indices of lexical cohesion showed statistically significant differences between texts and non-texts in the six genres, with emphasis on the Average Global Backward Vertex Cohesion Index, which is affected by text size. The global backward vertex and edge cohesion indices exhibited distinct behaviors per genre in their evolution throughout the text. Text classification using the Average Global Backward Vertex Cohesion Index and the logarithm of the number of periods achieved an average accuracy of 70%. It was concluded that network-based indices capture lexical repetition patterns that characterize texts, differentiating them from non-texts and capturing differences between genres, mainly the Average Global Backward Vertex Cohesion Index. The model contributes to unifying computational methods of textual analysis and psycholinguistic theories of text, formalizing its computational representation as a complex system. Based on the results, six hypotheses were constructed to expand the model.