Aplicação de Tecnologia Matching para realizar a Correspondência entre os Usuários e Especialistas na Prestação de Serviços de Assessoria Científica

• Ano de defesa: 2025
• Autor(a) principal: Souza, Elielson Nogueira de
• Orientador(a): Spanhol, Fabio Alexandre
• Banca de defesa: Andrade, Sidgley Camargo de, Brun, André Luiz, Silva, Thiago Henrique Pereira da
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Estadual do Oeste do Paraná; Cascavel
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação
• Departamento: Centro de Ciências Exatas e Tecnológicas
• País: Brasil
• Palavras-chave em Português: Pareamento; Processamento de linguagem natural; Consultoria científica
• Palavras-chave em Inglês: Pairing; Natural language processing; Scientific consulting
• Área do conhecimento CNPq: CIENCIA DA COMPUTAÇÃO
• Link de acesso: https://tede.unioeste.br/handle/tede/8145

Recent advances in Natural Language Processing (NLP), especially through the use of Deep Neural Networks, have promoted significant improvements in the interaction between humans and machines. These developments enable computers to understand and respond to human language with high accuracy, driving automation and operational efficiency in public and private institutions. This study aimed to develop an NLP-based matching model focused on achieving precise correspondence between users and scientific experts, applied within a platform specialized in academic review and scientific advisory services. The experiments were conducted using real-world data, applying cosine similarity between embedding vectors to identify the most relevant matches. The validation of results was performed using metrics such as Precision, Mean Reciprocal Rank (MRR), Normalized Discounted Cumulative Gain (nDCG), and Top-k Hit Ratio, in addition to comparison with a ground truth manually established by the SciBees technical team. This approach enabled a precise assessment of the models’ ability to retrieve relevant matches. The Bertimbau base and large versions of the models, as well as the OpenAI small and large versions, were tested. Of the models tested, the OpenAI-large model performed best, achieving average precision of 0.5667 and MRR of 0.8833, as well as a 100% hit ratio in the top three. This means that in all cases evaluated, at least one relevant expert was among the top three recommended. These results demonstrate the precision and consistency of the generated recommendations. Additionally, the methodology showed flexibility and replication potential, reinforcing its applicability in various domains such as healthcare, human resources, and agriculture.