FinBERTとRoBERTaトランスフォーマーモデルを用いた説明可能なESGセンチメント分析システム：財務リスクとビジネスインテリジェンスへの示唆
An Explainable ESG Sentiment Analysis System Using FinBERT and RoBERTa Transformer Models: Implications for Financial Risk and Business Intelligence (原題)
Ispita Jahan, Sonia Nashid, Kaniz Sultana Chy, Md Ashiqul Islam, M. Miah, Md Asif Hasan, Quazi Taif Sadat
🤖 gxceed AI 要約
日本語
FinBERTとRoBERTaにSHAP・XLexの説明機構を組み合わせ、ESG開示の感情分類を解釈可能にした研究。米国上場150社・1,200件の開示（2018–2024）を分析し、ESGセンチメントと財務リスク指標の関係を固定効果パネル回帰とグレンジャー因果で検証。ポジティブなESGセンチメントと市場ボラティリティに有意な逆相関（R²=0.67）を確認し、将来リスクへの増分予測価値も示した。
English
This study builds an explainable ESG sentiment model combining FinBERT and RoBERTa with SHAP and XLex interpretability. Analyzing 1,200 disclosures from 150 U.S. listed firms (2018–2024), it links ESG sentiment to financial risk via fixed-effects panel regression and Granger causality. Results show a significant inverse relationship between positive ESG sentiment and market volatility (R²=0.67), with incremental predictive value for future risk.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
SSBJ・有報でのESG開示が拡大する日本では、開示テキストを監査可能な形でリスク指標へ変換する手法への関心が高い。説明可能性を備えた本手法は、国内企業の統合報告書分析や投資家のESG評価高度化に示唆を与える。
In the global GX context
As ISSB/CSRD and SEC climate disclosure raise the volume of ESG text, explainable NLP pipelines that convert disclosures into auditable risk signals speak directly to global assurance and investor-trust debates. The paper's transparency focus addresses a key gap in opaque ESG ratings.
👥 読者別の含意
🔬研究者:Transformerと説明可能AIを組み合わせたESG感情分析の実証設計と因果検証手法を参考にできる。
🏢実務担当者:自社ESG開示がリスク指標にどう反映されるかを可視化する分析基盤の設計に活用できる。
🏛政策担当者:ESG開示のテキスト分析を監査・監督に組み込む際の説明可能性要件の議論に資する。
📄 Abstract（原文）
Environmental, Social, and Governance (ESG) disclosures have become a vital information source for supporting risk assessment and strategic decision-making in business intelligence systems. However, existing ESG analytics solutions often rely on opaque, accuracy-driven machine learning models that limit interpretability, auditability, and managerial trust. This study develops a transformer-based explainable ESG sentiment analysis model that integrates business intelligence principles with advanced natural language processing to support transparent and reliable financial risk assessment across U.S. industrial sectors. The proposed system combines FinBERT and RoBERTa transformer models with SHAP and XLex-based explainability mechanisms to support interpretable ESG sentiment classification. A corpus of 1,200 ESG-related disclosures from 150 publicly listed U.S. firms (2018–2024), including annual filings, sustainability reports, and corporate communications, is analyzed. ESG sentiment scores are aggregated at firm and sector levels and examined using fixed-effects panel regression and Granger causality analysis to evaluate their relationship with financial risk indicators, including market volatility, credit spreads, and investor confidence. Empirical results show a statistically significant inverse relationship between positive market volatility and ESG sentiment (R² = 0.67, p < 0.01), reflecting that constructive sustainability narratives contribute to financial stability. Granger causality tests further highlight that ESG sentiment provides incremental predictive value for future financial risk beyond historical risk measures. Explainability analyses demonstrate the dominant linguistic tokens and ESG dimensions driving model predictions, enhancing decision traceability and transparency. This study contributes to business intelligence and information systems engineering research by exploring an explainable ESG analytics framework that bridges advanced transformer models with interpretable decision support. The findings provide practical value for investors, managers, and regulators seeking transparent, reliable, and data-driven systems to translate ESG disclosures into actionable financial risk intelligence.
🔗 Provenance — このレコードを発見したソース
- semanticscholar https://www.wr-publishing.org/index.php/ijada/article/download/991/842first seen 2026-10-25 05:45:09 · last seen 2026-11-05 06:17:38
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録（無料・3キーワードまで）。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。