Scholarlink

<p>ESG Insight is an innovative framework designed for the automated extraction and analysis of Environmental, Social, and Governance (ESG) data from corporate disclosures, addressing the demand for accurate and standardized ESG information retrieval. Integrating Large Language Models (LLMs) with Retrieval Augmented Generation (RAG), the framework comprises an ESG metadata module, a report preprocessing unit, a Multi-Agent system, and a Multimodal Processing component. Its performance was evaluated on a sample of 300 SSE-listed companies’ ESG reports from year 2023 and 2024 across nine industries. Utilizing DeepSeek, ESG Insight achieved 78.2% accuracy in quantitative data extraction and 85.1% in disclosure identification, significantly surpassing baseline models. The analysis revealed disclosure rates of 67.8% for environmental, 55.9% for social, and 72.3% for governance indicators, underscoring persistent transparency gaps, particularly in social metrics. Limitations include challenges in processing visual data, such as charts, with planned enhancements to the Multimodal Processing module. Future research should focus on developing industry-specific datasets and extending the framework’s application to policy documents. ESG Insight enhances ESG data analysis, enabling stakeholders, securities firms and regulatory authorities to evaluate corporate sustainability, promote transparency, and advance China’s sustainable development objectives in alignment with SSE guidelines.</p>

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

ESG insight: a structured data extraction and evaluation framework via LLM

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Yuxie Zhou,
  • Xintao Wu,
  • Xin Xu,
  • Weixia Xu

摘要

ESG Insight is an innovative framework designed for the automated extraction and analysis of Environmental, Social, and Governance (ESG) data from corporate disclosures, addressing the demand for accurate and standardized ESG information retrieval. Integrating Large Language Models (LLMs) with Retrieval Augmented Generation (RAG), the framework comprises an ESG metadata module, a report preprocessing unit, a Multi-Agent system, and a Multimodal Processing component. Its performance was evaluated on a sample of 300 SSE-listed companies’ ESG reports from year 2023 and 2024 across nine industries. Utilizing DeepSeek, ESG Insight achieved 78.2% accuracy in quantitative data extraction and 85.1% in disclosure identification, significantly surpassing baseline models. The analysis revealed disclosure rates of 67.8% for environmental, 55.9% for social, and 72.3% for governance indicators, underscoring persistent transparency gaps, particularly in social metrics. Limitations include challenges in processing visual data, such as charts, with planned enhancements to the Multimodal Processing module. Future research should focus on developing industry-specific datasets and extending the framework’s application to policy documents. ESG Insight enhances ESG data analysis, enabling stakeholders, securities firms and regulatory authorities to evaluate corporate sustainability, promote transparency, and advance China’s sustainable development objectives in alignment with SSE guidelines.