-
Background This paper discusses the current state of Environmental, Social, and Governance (ESG) reporting, highlighting that companies use ESG reports to demonstrate their initiatives and commitments in these areas. ESG reports quantify corporate transparency and are crucial for evaluating a company's performance in ESG aspects, which is essential for investors and other stakeholders' decision-making. However, the vast number and varied formats of these reports pose significant challenges in integrating information at the corporate or industry level.
-
Existing Work Despite third-party ratings like those from MSCI, Sustainalytics, and Bloomberg, present research shows that no public ESG disclosure database with comprehensive metrics exists, which to some extent limits the depth of analysis and regulation of corporate ESG performance.
- Introduced ESGReveal
-
Challenge 1: Systematic extraction of key data from ESG reports The ESGReveal method overcomes integration challenges by integrating advanced Large Language Models (LLM) with Retrieval Augmented Generation (RAG) techniques to ensure consistent and accurate retrieval of ESG information. It was used to evaluate data extraction effectiveness from ESG reports of companies across 12 industries listed on the Hong Kong Stock Exchange. The results showed GPT-4 achieving 76.9% accuracy in data extraction and 83.7% in disclosure analysis, surpassing baseline models.
-
Challenge 2: Comparative performance of multiple models The paper identifies the variability in the analytical performance of LLMs and proposes more research to improve and compare the analytical performance of various LLMs, addressing issues of inadequate data disclosure and lack of transparency.
-
ESGReveal implementation includes three modules: ESG metadata module, report preprocessing module, and LLM agent module. The study involved data from a representative sample of 166 companies based on industry distribution and market capitalization. Empirical results show that the development of ESGReveal demonstrates the potential of large language models to improve the accuracy of ESG data analysis and provides tools for evaluating and enhancing corporate sustainability practices. It contributes to advancing the transparency of corporate reporting and supports the achievement of sustainable development goals.
ESGReveal marks significant progress in ESG data processing, aiming to improve the consistency and accuracy of structured data extraction from corporate reports through large language models and related techniques, and it has driven improvements in ESG practices and transparency.