Vector representation of texts applied to prediction models

• Ano de defesa: 2020
• Autor(a) principal: Stern, Deborah Bassi
• Orientador(a): Izbicki, Rafael
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa Interinstitucional de Pós-Graduação em Estatística - PIPGEs
• Palavras-chave em Português: Processamento de linguagem natural; Redes neurais; Representação vetorial de palavras; Modelos de predição
• Palavras-chave em Inglês: Natural language processing; Neural networks; WordVectors; Prediction models
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::PROBABILIDADE E ESTATISTICA
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/12362

Natural Language Processing has gone through substantial changes over time. It was only recently that statistical approaches started receiving attention. The Word2Vec model is one of these. It is a shallow neural network designed to fit vectorial representations of words according to their syntactic and semantic values. The word embeddings acquired by this method are state-of-art. This method has many uses, one of which is the fitting of prediction models based on texts. It is common in the literature for a text to be represented as the mean of its word embeddings. The resulting vector is then used in the predictive model as an explanatory variables. In this dissertation, we propose getting more information of text by adding other summary statistics besides the mean, such as other moments and quantiles. The improvement of the prediction models is studied in real datasets.