A carregar...
Publicação
Estudo da interligação das intenções de voto, através da adaptação de um modelo de linguagem natural
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 10.47 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Ao longo dos últimos anos, as eleições portuguesas têm sido cada vez mais acompanhadas
da realização de inquéritos e sondagens de intenções de voto. Com os resultados
inesperados das eleições autárquicas de 2021 e eleições legislativas de 2022, surgem também
as dúvidas em relação à eficácia das previsões efetuadas.
Os espaços onde se aplicam técnicas de estatística apresentam ligações que comprometem
a premissa de independência dos objetos em estudo. Uma das formas de contornar
de espaços de elevada dimensão é através da projeção dos mesmos num espaço finito e
bem definido matematicamente.
Foi aplicado modelo Word2Vec, em conjunto com o método de decomposição de valores
singulares, a um espaço de sondagens, realizando uma projeção do mesmo a um
espaço de Hilbert. Foi feita uma análise de diversas variáveis do modelo, onde se utilizou
a data da sondagem, a numeração dos inquiridos e a respetiva intenção de voto, com o
objetivo de construir a base de um modelo capaz de representar o espaço de sondagens e
as suas ligações.
Pela análise gráfica das projeções do espaço de intenções de voto e respetivos gráficos
de valores singulares, foi possível concluir que o melhor modelo consiste no uso
de um Corpus com uma única frase, uma camada oculta de 5 dimensões e do uso de
uma janela de contexto com todos os vocábulos. Observou-se ainda que, ao longo dos
momentos temporais considerados, as perdas de informação que ocorrem nas reduções
de dimensionalidade diferem para todos os partidos.
Over the last few years, the portuguese political elections have been increasingly accompanied by surveys and polls of voting intentions. With the surprising outcomes of the 2021 municipal and 2022 legislative elections, the accuracy of the forecasts made in the electoral polls is becoming another source of debate. Many of the mathematical spaces where statistical techniques are applied have connections that compromise the premise of independence of objects under study. One of the solutions to get around high-dimensional spaces is by projecting them into a finite and mathematically well-defined space. TheWord2Vec model was applied, together with the singular value decomposition, to an electoral polls space, performing a projection of it in a Hilbert’s Space. A multivariate study of the model was conducted, where the only data used was the poll respondents’ numbers and the corresponding vote intentions, with the purpose of creating the framework for a model that can depict the electoral polls space and its connections. The best model, according to graphical analysis of the space projections and corresponding graphs of singular values, uses a Corpus containing a single sentence, a 5-dimensional hidden layer, and a context window with all the words. Additionally, it was noted that the information losses brought on by the dimensionality reductions vary for each party over the temporal moments taken into account.
Over the last few years, the portuguese political elections have been increasingly accompanied by surveys and polls of voting intentions. With the surprising outcomes of the 2021 municipal and 2022 legislative elections, the accuracy of the forecasts made in the electoral polls is becoming another source of debate. Many of the mathematical spaces where statistical techniques are applied have connections that compromise the premise of independence of objects under study. One of the solutions to get around high-dimensional spaces is by projecting them into a finite and mathematically well-defined space. TheWord2Vec model was applied, together with the singular value decomposition, to an electoral polls space, performing a projection of it in a Hilbert’s Space. A multivariate study of the model was conducted, where the only data used was the poll respondents’ numbers and the corresponding vote intentions, with the purpose of creating the framework for a model that can depict the electoral polls space and its connections. The best model, according to graphical analysis of the space projections and corresponding graphs of singular values, uses a Corpus containing a single sentence, a 5-dimensional hidden layer, and a context window with all the words. Additionally, it was noted that the information losses brought on by the dimensionality reductions vary for each party over the temporal moments taken into account.
Descrição
Palavras-chave
Word2Vec Decomposição de valores singulares Machine learning Sistema em expansão