企業永續報告之大型語言模型輔助地理分析： 以臺灣電池產業所在之環境風險揭露為例

Abstract

隨著近年來氣候變遷極端天氣事件發生的頻率增加，氣候變遷引發企業經營的挑戰與風險，受到企業利害關係人極大的關注。為此，永續報告書的揭露內容開始規範企業必須揭露營運區域面臨的氣候變遷風險和因應策略。我國行政院金融監督管理委員會亦規範上市資本額高於20億之高能耗產業（如本研究關注的電池業），需加強永續揭露指標內容。然而當前永續報告書之分析或評比方式，甚少根據企業營運所在地、乃至整體產業鏈之空間分布型態，揭露相關環境特性及氣候變遷的風險暴露。 本研究以臺灣電池產業相關企業57本英文永續報告書為樣本；採用大型語言模型（Large-Language Model, LLM） GPT‑4o 與 Llama‑3.1‑8B，以檢索強化生成（Retrieval-Augmented Generation, RAG）設計輔助系統，組織自動辨識地理實體、極端天氣事件的分析流程，並依據企業營運據點實際暴露的災害潛勢資訊比對報告書環境風險揭露內容，探討 LLM 產製的報告書品質提升建議之可行性。研究結果發現當前在環境相關揭露內容中，目前各家企業揭露內容仍有受法規引導的現象存在，且地理資訊與極端天氣事件資訊篇幅極少，使得 RAG 系統設計評估任務時，在檢索目標文本上備受挑戰。另外當前國內業者偏好以圖表彙整資訊，然這對於需要仰賴連貫上下文判斷詞彙語義的 LLM 來說，如何將圖表資訊轉化為 LLM 可理解之型態，將是永續報告書自動化評估或生成任務的關鍵。在解讀模型成果方面，引入 F1- Score 能從混淆矩陣探討哪些生成結果其實是模型幻覺，然而 F1-Score 指標仍仰賴研究者事先建立評估資料集，前置作業暫無自動化流程，建議未來研究者可更深入發展語言模型生成結果評估的各項方法。

With the increasing frequency of extreme weather events in recent years, the challenges and risks posed by climate change to business operations have garnered significant attention from corporate stakeholders. In response, sustainability reporting standards now mandate that companies disclose the climate-change-related risks they face in their operational regions and their corresponding mitigation strategies. Taiwan's Financial Supervisory Commission (FSC) has also stipulated enhanced sustainability disclosure indicators for high-energy-consumption industries, such as the battery sector focused on in this study, specifically targeting listed companies with a paid-in capital of over NT$2 billion. However, current analyses and evaluations of sustainability reports seldom incorporate the spatial distribution of a company's operations, let alone that of its entire supply chain, to disclose relevant environmental characteristics and exposure to climate change risks. This study analyzes a sample of 57 English sustainability reports from companies in Taiwan's battery industry. We employ the Large Language Models (LLMs) GPT-4o and Llama-3.1-8B, utilizing a Retrieval-Augmented Generation (RAG) framework to design an auxiliary system. This system facilitates an analytical workflow that automatically identifies geographic entities and extreme weather events. The environmental risk disclosures within the reports are then compared against disaster potential data corresponding to the actual locations of the companies' operating sites. The research explores the feasibility of using LLMs to generate recommendations for enhancing the quality of these reports. Our findings indicate that current environmental disclosures are heavily influenced by regulatory guidance, with minimal content dedicated to geographic information and extreme weather events. This scarcity of specific textual information presents a significant challenge for the retrieval task within the RAG system. Furthermore, companies in Taiwan currently favor the use of charts and tables to summarize information. For LLMs, which rely on continuous contextual data to interpret semantic meaning, a key challenge for the automated evaluation or generation of sustainability reports will be the transformation of this graphical information into a format that is machine-readable. In terms of interpreting model outputs, the introduction of the F1-Score, derived from a confusion matrix, allows for an examination of which generated results may be model hallucinations. Nevertheless, the F1-Score metric still depends on the prior creation of an evaluation dataset by researchers, a preparatory task for which there is currently no automated process. Future research should therefore focus on further developing various methods for evaluating the outputs of language models.