Scholarlink

The United Nations (UN) has developed the 2030 Agenda, consisting of 17 Sustainable Development Goals (SDGs) aimed at ensuring the planet’s sustainability and humanity’s well-being. However, tracking scientific progress towards these goals poses a challenge for government agencies. To address this, we propose using a language model capable of automatically classifying texts into one of the 17 SDGs. We used an imbalanced dataset of scientific paper abstracts previously classified into these categories and our approach involved fine-tuning a multi-class sequence classification model using Google’s Bidirectional Encoder Representations from Transformers (BERT) with a classification head. Our model’s average accuracy is 88.61% and over 70% class-wise precision for most SDGs. As a use case, we also classify research projects submitted to FACEPE, the research funding agency of Pernambuco, Brazil. We present the results and discuss limitations and potential future steps.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Classification of UN’s Sustainable Development Goals Using Bidirectional Encoder Representations from Transformers Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Leonides Medeiros Neto,
  • Maicon Herverton Lino Ferreira da Silva Barros,
  • Raysa Carla Leal da Silva,
  • Roberto Cesar da Silva Leal,
  • Guto Leoni Santos,
  • Theo Lynn,
  • Raphael Augusto de Sousa Dourado,
  • Patricia Takako Endo

摘要

The United Nations (UN) has developed the 2030 Agenda, consisting of 17 Sustainable Development Goals (SDGs) aimed at ensuring the planet’s sustainability and humanity’s well-being. However, tracking scientific progress towards these goals poses a challenge for government agencies. To address this, we propose using a language model capable of automatically classifying texts into one of the 17 SDGs. We used an imbalanced dataset of scientific paper abstracts previously classified into these categories and our approach involved fine-tuning a multi-class sequence classification model using Google’s Bidirectional Encoder Representations from Transformers (BERT) with a classification head. Our model’s average accuracy is 88.61% and over 70% class-wise precision for most SDGs. As a use case, we also classify research projects submitted to FACEPE, the research funding agency of Pernambuco, Brazil. We present the results and discuss limitations and potential future steps.