Uma solução baseada em template e multi-solução para geração de textos a partir de triplas

• Ano de defesa: 2020
• Autor(a) principal: Mota, Abelardo Vieira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Geração de linguagem natural; Bases de conhecimento; Geração dados-para-texto
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/53756

With the evolution of studies on the Semantic Web, the RDF data representation format was proposed and several knowledge bases were created with it. As a result, a lot of information has become available for computer processing and for expert users, but at the same time this information is inaccessible to a lay user. One way to make this type of data more accessible is to represent it using natural language, which can be costly if done manually. This context motivates the development of solutions that automatically generate texts from RDF databases. There are several integrated solutions proposed for this purpose, but they require large, high-quality databases for training, they lack controls over how the text will be generated and do not offer guarantees that the generated text verbalizes all and only the input data. This work addresses these problems by proposing a modular template-based solution based and with a multiple solution approach, which generates multiple texts during the generation process. The use of templates requires an amount of data proportional to the variety of texts to be generated, and the modularization of the solution allows greater control over the generation process, as well the possibility of giving greater guarantees over the final result. The experiments carried out to validate the proposed solution used a real database and demonstrate that the proposed approach generates better quality texts and outperforms some state-of-the-art solutions, regarding the BLEU, METEOR and TER evaluation methods. In addition to the evaluation with automatic methods, an inspection was also carried out on a sample of the generated texts in order to highlight limitations and possible improvement points.