Using complex networks and Deep Learning to model and learn context

• Ano de defesa: 2020
• Autor(a) principal: Júnior, Edilson Anselmo Corrêa
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Ambiguidade; Ambiguity; Complex networks; Context; Contexto; Deep learning; Redes complexas
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-16022021-151616/

The structure of language is strongly influenced by the context, whether it is the social setting, of discourse (spoken and written) or the context of words itself. This fact allowed the creation of several techniques of Natural Language Processing (NLP) that take advantage of this information to tackle a myriad of tasks, including machine translation, summarization and classification of texts. However, in most of these applications, the context has been approached only as a source of information and not as an element to be explored and modeled. In this thesis, we explore the context on a deeper level, bringing new representations and methodologies. Throughout the thesis, we considered context as an important element that must be modeled in order to better perform NLP tasks. We demonstrated how complex networks can be used both to represent and learn context information while performing word sense disambiguation. In addition, we proposed a context modeling approach that combines word embeddings and a network representation, this approach allowed the induction of senses in an unsupervised way using community detection methods. Using this representation we further explored its application in text classification, we expanded the approach to allow the extraction of text features based on the semantic flow, which were later used in a supervised classifier trained to discriminate texts by genre and publication date. The studies carried out in this thesis demonstrate that context modeling is important given the interdependence between language and context, and that it can bring benefits for different NLP tasks. The framework proposed, both for modeling and textual feature extraction can be further used to explore other aspects and mechanisms of language.