Modelo neural recozido para a representação semântica de documentos por meio de vetores contínuos

• Autor(a) principal: Mendonça, Leandro Rezende Carneiro de
• Data de Publicação: 2020
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Repositório Institucional da UFG
• Texto Completo: http://repositorio.bc.ufg.br/tede/handle/tede/10965
• topic: Representação de documento; Redes neurais; Processamento de linguagem natural; Análise de texto; Representação vetorial; Otimização; Recozimento simulado; Aprendizado de máquina; Document representation; Neural network; Natural language process; Text analysis; Vector representation; Optimization; Simulated annealing; Machine learning; ENGENHARIAS

As a result of the growing production of unstructured textual data, techniques for representing words and documents in the vector space have emerged recently. The Brazilian Public Ministry has received several textual requests that are send by citizens with different needs, such as those involved in cases of domestic violence against women, others requesting intensive care unit admissions, and more. The time spent in classifying, detecting similar requests and distributing them is essential to optimize and save public resources. Therefore, we adopted the neural model with the Simulated Annealing (SA), a classic global optimization algorithm with low computational complexity, because of the need to reduce the daily training time, providing a more friendly graphic visualization of data in high dimensions, supporting the judicial decision process. The physical analogy of the SA meta-heuristic associated with the continuous representation of documents in the vector space contribute greatly to the friendly visualization of a high-dimensional dataset, maintaining a comparable score with other deep models and optimization algorithms, such as Covariance Matrix Adaptation Evolution Strategy (CMA-ES) and Bayesian Optimization (BO).