ESG文本具有財務資訊性嗎？大型語言模型摘要與情緒的角色探討

Abstract

本研究旨在探討企業永續報告書中關於環境、社會與公司治理（ESG）揭露內容是否具有財務資訊價值，並進一步分析大型語言模型（Large Language Models, LLMs）所生成之摘要文本與其所包含之情緒對企業財務績效與市場評價的影響。本文蒐集2010年至2024年間標準普爾500指數（S&P 500）中非金融類企業之ESG報告書，運用Gemini 1.5 Flash模型進行報告摘要生成，並透過FinBERT模型進行情緒分析。實證結果顯示，相較於原始報告書文本所萃取之情緒，LLM生成摘要中所反映之情緒與企業當期與次期資產報酬率（ROA），以及次期Tobin’s Q，皆呈現顯著正向關聯，顯示LLM摘要具備更高之資訊性與預測性。此外，傳統機構提供之ESG評等在本研究中未能展現顯著解釋力。綜合而言，本文證實大型語言模型可有效濃縮冗長的永續揭露內容，提升其可讀性與財務相關性，為資本市場參與者提供具決策意義之永續資訊，並為ESG文本分析提供嶄新的研究視角與方法論貢獻。

This study examines whether the narrative content of Environmental, Social, and Governance (ESG) reports contains financially informative signals, and further investigates the role of summary sentiment generated by Large Language Models (LLMs) in relation to firm performance and valuation. Utilizing a sample of ESG reports from non-financial firms listed on the S&P 500 between 2010 and 2024, we apply the Gemini 1.5 Flash model to generate concise summaries, followed by sentiment analysis using the FinBERT model. Empirical results reveal that, compared to sentiments derived from full-length ESG texts, sentiments extracted from LLM-generated summaries exhibit a significantly positive association with both contemporaneous and subsequent return on assets (ROA), as well as future Tobin’s Q. These findings suggest that LLM summaries enhance the informativeness and predictive power of ESG disclosures. In contrast, institutional ESG ratings do not demonstrate significant explanatory power in the same models. Overall, this study highlights the potential of LLMs to distill value-relevant content from lengthy ESG narratives, improve information processing efficiency, and enhance the decision-usefulness of sustainability disclosures. The findings contribute to the growing literature on AI-assisted financial text analysis and offer a novel methodological approach to ESG research.