| 摘要: | 隨著永續發展概念的興起，環境、社會與公司治理(ESG)評分已成為投資者評估企業永續表現及投資價值的重要依據。然而目前的ESG評分主要由第三方評鑑機構根據企業自願揭露的報告、政策與財務數據進行評估，可能存在資訊不對稱或評估標準不一致的問題。因此如何利用公開資訊來提高ESG評分的透明度與即時性，成為企業與投資人關注的重點。本研究聚焦於台灣前百大上市櫃公司，探討公開ESG新聞與ESG評分之間的關聯性。本研究選擇來自第三方的ESG新聞作為訓練資料，以確保資訊來源的中立性與客觀性，並開發一套專為台灣企業設計的繁體中文新聞分析模型，以預測企業的ESG評分。研究數據主要來自臺灣經濟新報社會責任新聞作為模型的自變數，並透過不同的詞嵌入方法(如Word Embedding、TF-IDF)進行文本特徵提取，最後結合多種機器學習與迴歸模型(如隨機森林、XGBoost、深度學習模型)進行預測，以找出最適合評估B級以上ESG評分的模型組合。研究結果顯示，ESG新聞與企業ESG評分之間存在顯著的相關性，且特定詞嵌入技術與迴歸模型的組合能有效提高預測準確度。本研究的貢獻在於提供一種基於新聞文本的ESG評分預測方法，為投資者、企業與監管機構提供更具即時性與透明度的評估工具，並進一步促進台灣企業對ESG議題的重視與資訊揭露的完善性。;With the rise of sustainable development concepts, Environmental, Social, and Governance (ESG) ratings have become a crucial criterion for investors to assess corporate sustainability performance and investment value. However, current ESG ratings are primarily evaluated by third-party rating agencies based on voluntarily disclosed corporate reports, policies, and financial data, which may result in information asymmetry or inconsistencies in evaluation standards. Therefore, how to leverage publicly available information to enhance the transparency and timeliness of ESG ratings has become a key concern for both corporations and investors.
This study focuses on Taiwan’s top 100 listed companies and explores the relationship between publicly available ESG news and ESG ratings. To ensure the neutrality and objectivity of information sources, ESG news from third-party sources is used as training data. A Traditional Chinese news analysis model tailored for Taiwanese enterprises is developed to predict corporate ESG ratings. The research data primarily originates from ESG-related news provided by the Taiwan Economic Journal, which serves as independent variables for the model. Various word embedding techniques, such as Word Embedding and TF-IDF, are applied for text feature extraction. Additionally, multiple machine learning and regression models, including Random Forest, XGBoost, and deep learning models, are employed to identify the most suitable model combination for evaluating ESG ratings of grade B and above.
The results indicate a significant correlation between ESG news and corporate ESG ratings. Furthermore, specific combinations of word embedding techniques and regression models effectively enhance prediction accuracy. The primary contribution of this study is the development of an ESG rating prediction method based on news text, providing investors, corporations, and regulatory bodies with a more timely and transparent assessment tool. This, in turn, promotes greater corporate awareness of ESG issues and improves the comprehensiveness of ESG information disclosure in Taiwan. |
