Estudo de modelos de word embedding

• Autor(a) principal: Sousa, Samanta de
• Data de Publicação: 2016
• Tipo de documento: Trabalho de conclusão de curso
• Idioma: por
• Título da fonte: Repositório Institucional da UTFPR (da Universidade Tecnológica Federal do Paraná (RIUT))
• Texto Completo: http://repositorio.utfpr.edu.br/jspui/handle/1/12522
• topic: Inteligencia Artificial; Processamento de linguagem natural (Computação); Bibliotecas digitais; Artificial intelligence; Natural language processing (Computer science); Digital libraries; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO

The area of Artificial Intelligence seeks to construct mechanisms that simulate the intelligence of the Human beings so that they perform tasks that help them. There is the field of Natural Language Processing, an AI sub-area that seeks to understand and To generate the natural language, in this way the PLN is used by AI as a means to The mechanisms that use the natural language in its execution, such as writing and production Of a text, translation, learning and teaching among others. The language follows a format Not difficult to process by the computer, such as sd morphological variations and Syntactic as well as the ambiguity in the natural language that hinder the process of comprehension, In this way, area methodologies convert such information so that the manipulation Computer are easier. Among the information representations Existing Word Embedding technique is currently in the PLN field, where The information is represented in vectors where their values are similar when the Words are similar, that is, it is a representation that encodes similarity relations Between the words besides having a low computational cost. In this way the goal of Work was to carry out a comparison between three models ofWord Embeddings Cbow, Skip- Gram and Glove with the purpose of identifying which presents better performance in the generation of Vectors of representation of words (embeddings). First, construction was carried out Of a corpus using Wikipedia in sequence, the pre-processing of those corpus Information to be used as a training set, the models were trained Using scripts that are created using the Gensim and Glove Python libraries, the Embedding evaluations were done with the files available from Pennington et al. (2014), where in each evaluation / test the parameters were modified in order to verify the Their influence on the performance of models. Some specific settings for running Of the training of the models were identified and reported in the study, the results obtained Demonstrated that the Cbow was the model that presented better performances in the majority Of the tests. It has been found that the Word Embeddings technique fairly Similarity information between words even with the values of the parameters being Small compared to other jobs.