CarbonCrediBERT：専門家誘導型フューショットおよび選択的分類による炭素クレジット開示の分析

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:AI×ESG開示分析における少数ラベル・クラス不均衡・ハードネガティブ設計の実証的ベンチマークを提供する。

🏢実務担当者:自社報告書のクレジット・オフセット記述を機械的に棚卸しし、開示品質やグリーンウォッシュリスクを点検する際に活用できる。

🏛政策担当者:カーボンクレジット主張の大規模モニタリング手法として、開示規制の執行・検証設計に示唆を与える。

<title>Abstract</title> <p>Corporate reports increasingly discuss carbon credits in relation to net-zero targets, offsetting claims, and climate transition plans, particularly as climate-related financial disclosure requirements expand. Analysing these disclosures at scale remains difficult because terms such as credit, offset, and net also occur in unrelated financial, accounting, and operational contexts. Expert-labelled data are also costly to produce and often contain substantial class imbalance. We propose CarbonCrediBERT, an expert-guided natural language processing framework that classifies paragraphs according to carbon credit relevance, sentiment, and reported credit use, while referring uncertain cases for human review. CarbonCrediBERT combines an expert defined taxonomy, hard-negative examples drawn from closely related but nonrelevant passages, and classification methods suited to small expert-labelled datasets. Using annual, climate, sustainability, and ESG reports published by ASX100 companies in 2024–2025, the framework produced 8,299 keyword-matched sentence records and a balanced candidate pool of 15,026 paragraphs. We evaluated ClimateBERT and DistilRoBERTa fine-tuning and SetFit–MiniLM few-shot learning using three task-specific datasets of 100 expert-labelled paragraphs each. Five-fold development cross-validation selected fine-tuned ClimateBERT for carbon credit relevance, achieving 93.8% accuracy, and SetFit–MiniLM for sentiment and reported credit use. These results show that broad climate-language detection alone is insufficient for identifying carbon credit disclosures and that no single modelling approach performs best across all three tasks. They also highlight the continuing importance of expert judgement, particularly for sentiment and ambiguous statements about credit use. Future work will expand the domain corpus and expert-labelled evidence and evaluate transfer across companies, reporting years, and jurisdictions, establishing a pathway towards a reusable CarbonCrediBERT foundation model for carbon credit disclosure analysis.</p>

• Research Square https://doi.org/10.21203/rs.3.rs-11679188/v1first seen 2026-11-13 04:28:55 · last seen 2026-11-14 04:22:44
• openalex https://doi.org/10.21203/rs.3.rs-11679188/v1first seen 2026-11-13 05:13:49 · last seen 2026-11-14 05:08:19