Deep Learning NER Pipeline for Automated Basel III / IFRS 9 Risk Parameter Extraction from Climate Narratives
気候関連ナラティブからの自動Basel III/IFRS9リスクパラメータ抽出のための深層学習NERパイプライン (AI 翻訳)
Rohit Nimmala, Jagrut Nimmala, Milan Parikh
🤖 gxceed AI 要約
日本語
本論文は、気候変動シナリオのテキストからBasel III/IFRS 9の信用リスクパラメータ（PD、LGD、EAD等）を自動抽出する深層学習パイプラインClimRiskNERを提案する。金融・気候の二重ドメイン適応事前学習と制約付き系列生成を組み合わせ、ECB、BoE、NGFSのストレステスト報告書等で高い性能（F1 0.82-0.89）を達成した。手動抽出と比較して99%以上の時間短縮を実現し、SHAPによる説明可能性も備える。
English
This paper introduces ClimRiskNER, a deep learning pipeline that automatically extracts Basel III/IFRS 9 credit risk parameters (PD, LGD, EAD) from climate scenario narratives. It combines dual-domain adaptive pre-training (financial and climate) with constrained seq2seq decoding, achieving F1 scores of 0.82-0.89 on ECB, BoE, and NGFS stress test reports and TCFD disclosures. The system reduces extraction time by over 99% versus manual review and provides SHAP-based provenance for audit trails.
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
日本でもSSBJの開示基準や有報での気候関連情報記載が進む中、本パイプラインは定性情報から定量的リスクパラメータを自動抽出する手法を提供する。日本の金融機関がTCFD開示やストレステストを効率化し、規制対応を強化する上で有用。
In the global GX context
This pipeline directly addresses the global need to integrate climate risks into financial risk management under ISSB, CSRD, and Basel III frameworks. By automating extraction of credit risk parameters from unstructured climate narratives, it enables scalable, auditable analysis for banks, regulators, and investors worldwide.
👥 読者別の含意
🔬研究者:For NLP and climate finance researchers, this work demonstrates a novel approach to structured information extraction from specialized domains, combining domain-adaptive pre-training and constrained decoding.
🏢実務担当者:Corporate sustainability and risk teams can adopt this pipeline to automate the extraction of climate risk parameters from disclosures, reducing manual effort and improving data consistency for reporting.
🏛政策担当者:Regulators and central banks can leverage this tool to systematically assess climate risk exposures across financial institutions and enforce disclosure standards more effectively.
📄 Abstract（原文）
Banks must manually extract credit risk parameters, including probability of default (PD), loss given default (LGD), exposure at default (EAD), and expected credit loss (ECL), from qualitative climate scenario narratives published by central banks and TCFD-reporting firms. Existing financial NER systems target generic entity types or XBRL tags, and climate NLP methods classify disclosures at the document level but do not extract quantitative risk parameters. We introduce ClimRiskNER, a novel end-to-end pipeline that combines dual-domain adaptive pre-training (financial and climate corpora) with constrained seq2seq decoding to extract structured Basel III/IFRS 9 parameters from unstructured climate text. The pipeline uses FinBERT-Climate, a FinBERT model further pre-trained on over 2 million climate paragraphs, for token-level NER, followed by a T5-based seq2seq module with prefix-trie constrained beam search enforcing output conformity to a Basel III/IFRS 9 ontology. On a curated dataset of ECB, BoE, and NGFS climate stress test reports and TCFD disclosures, the pipeline achieves entity-level F1 of 0.89 for PD, 0.85 for LGD, and 0.82 for EAD, outperforming the next-best baseline (GPT-4 few-shot) by 9 F1 points. The system reduces extraction time by over 99% versus manual review (0.3 seconds per page versus 720 seconds) and provides SHAP-based token-level provenance for regulatory audit trails.
🔗 Provenance — このレコードを発見したソース
- semanticscholar https://doi.org/10.1109/icaiset66439.2026.11542084first seen 2026-06-05 05:22:19
🔔 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。