Spring til hovednavigation Spring til søgning Spring til hovedindhold

Bridging the ESG Data Gap: Classifying Textual Data for SDGs by Leveraging Natural Language Processing

Guangfei Feng & Luxin Wang

Studenteropgave: Kandidatafhandlinger

Abstract

As ESG continues to be a universal topic for policymakers, investors, and companies, challenges remain in increasing transparency and timeliness of ESG-related data. This has prompted the need to seek alternative data to better understand a company's ESG performance and progress towards ESG initiatives. To effectively identify ESG issues from different sources of data, this thesis explores the application of machine learning models for the multi-label classification of text data related to the Sustainable Development Goals (SDGs) - a widely adopted ESG framework. Additionally, models for multi-class classification are used for comparison. Specifically, four models were trained for the two types of tasks: BERT, BERT Large, word2vec and bag of words, and dummy classifier as the baseline model. Three datasets were used for training each model: a single-labeled-only dataset, a multi-labeled-only dataset, and a mixed dataset. The objective is to identify the most effective model for SDGs text classification and to evaluate the impact of different training data on different models. The study found that BERT large trained with multi-labeled-only dataset provided the best results for SDGs multi-label classification. However, all the models for multi-class classification outperformed models for multi-label classification. The study also provides insights into the challenges and opportunities of SDGs text classification, including the need for reliable data for model training, as well as the potential for using the models to support ESG reporting and monitoring. Overall, this thesis contributes to ESG topic classification research by comparing natural language processing models using different datasets and establishes a foundation for future work.

UddannelserCand.merc.it Business Administration and Information Systems, (Kandidatuddannelse) Afsluttende afhandling
SprogEngelsk
Udgivelsesdato15 maj 2023
Antal sider113
VejledereDaniel Hardt