A deep learning model for identifying academic publications aligned with the sustainable development goals

• Ano de defesa: 2023
• Autor(a) principal: Dias, Alexandre Henrique Soares
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Rio Grande do Norte; Brasil; UFRN; PROGRAMA DE PÓS-GRADUAÇÃO EM ENGENHARIA ELÉTRICA E DE COMPUTAÇÃO
• Palavras-chave em Português: Sustainable development; Multilabel classification; Deep learning; Natural language; Scientometrics; CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: https://repositorio.ufrn.br/handle/123456789/55191

In 2015, the United Nations established the 17 Sustainable Development Goals (SDGs) to promote environmental stewardship, economic advancement, and social equity. Within this framework, scientific research plays a pivotal role in addressing the challenges encompassed by the SDGs. This work presents the development and implementation of a multi-label classifier for identifying scientific articles aligned with the SDGs, established by the United Nations. Recognizing the inherent complexities and challenges associated with the SDGs, the work aims to assist governmental, educational, and private institutions in making informed decisions related to SDG implementation. The proposed classifier was trained and evaluated on a dataset of approximately one million samples from the Scopus database, consisting of titles of scientific publications and SDG labels. Three distinct deep learning architectures were explored: Recurrent Neural Networks, BERT, and Distilled BERT. Special consideration was given to handling class imbalance, and a new multi-objective metric, F-Green, was proposed as a way to assess the models’ performances considering not only their performance metrics but also their carbon footprint. Among the model architectures experimented with, the Distilled BERT architecture was found to provide the optimal balance between performance and environmental impact. A proof-of-concept web application was developed to demonstrate the model’s functionality, allowing users to interact with the model and classify hypothetical academic article titles according to one or more SDGs. The work highlights the importance of aligning scientific production with sustainability goals and provides a practical tool to support decision-making in academia and beyond.