Comparing vector document representation methods for authorship identification

• Ano de defesa: 2021
• Autor(a) principal: Quintanilla, Pamela Rosy Revuelta
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado máquina; Atribuição de autoria; Authorship attribution; Classificação de texto; Complex networks; Extração de características; Feature extraction; Graph embedding; Graph embeddings; Machine Learning; Redes complexas; Text classification; Word embeddings
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/45/45134/tde-05052021-040638/

Over the years the information available in online media has had a great increase. In this sense, the automation of processing languages natural for large amounts of information gained importance, for example, text classification task. It can be used to identify the author (Authorship Identification); however, it requires Machine Learning techniques to identify the author, these techniques have given good results in NLP. In addition, Machine Learning receives the feature vector of the texts, which is extracted using vector document representation methods. The methods proposed for this research are grouped into three different approaches: i) methods based on vector space models, ii) methods based on word embeddings, and iii) methods based on graph embeddings, for this approach, we first model the texts as graphs. On the other hand, not all the methods are used for different languages because they can have different efficiency depending on the language of the analyzed texts. Therefore, the objective of this research is to compare several of these methods using literary texts in English and Spanish. In this way, we analyze whether the methods are efficient to represent several languages or its performance depends on the characteristic of every language. The results showed that the methods of Graph embeddings achieved the best performance for both languages, being that English reached a fairly high success rate. On the other hand, the other methods achieved good performance for English, however, the results for Spanish were not optimal. We believe that the results in Spanish were worse due to the morphological, lexical, and syntactic complexity that this language presents in comparison to English. For this reason, different approaches were compared for the mathematical representation of texts that try to cover the different aspects of a language.