Aplicação de programação genética na análise de sentimentos

• Ano de defesa: 2018
• Autor(a) principal: Bordin Junior, Airton
• Orientador(a): Silva, Nádia Félix Felipe da
• Banca de defesa: Silva, Nádia Félix Felipe da, Camilo Junior, Celso Gonçalves, Rosa, Thierson Couto, Covões, Thiago Ferreira, Fernandes, Deborah Silva Alves
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Ciência da Computação (INF)
• Departamento: Instituto de Informática - INF (RG)
• País: Brasil
• Palavras-chave em Português: Análise de sentimentos; Mineração de opiniões; Programação genética; Classificadores
• Palavras-chave em Inglês: Sentiment analysis; Opinion mining; Genetic programming; Classifiers
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/9211

The Web is commonly used as a platform for debates, opinions, evaluations, etc. These data allowed the area of Sentiment Analysis (SA) to develop to extract information and knowledge that can be used in different applications. Among the challenges of SA we can highlight the creation of classifiers with good efficacy. Typically, the classification models are generated using specific heuristics, manually defined and not adaptable to different contexts. Thus, this work proposes the automated generation of hybrid SA classifiers - with Machine Learning (ML) techniques and lexical dictionaries - using Genetic Programming (GP). It is expected to reduce the cost of generating the classifiers and increase the predictive power for each domain analyzed. The goal is that these classifiers will be competitive with the classical ML algorithms used in SA, generalizable, adaptable to the context and able to determine the relevance of each lexical to the applied domain. In addition, the aim is allow to aggregate other ML techniques to create even more effective hybrid solutions. In order to validate the proposal, SemEval 2014 benchmark was used. The results show that the approach with GP is promising since the generated models are competitive, and sometimes better, with other researches. The ensemble proved to be effective in increasing the predictive power of the system, obtaining better results than the use of the techniques individually. Finally, we highlight the ability of models customization according to the context approached and the possibility of knowledge transfer of the users through the functions used by GP.