グリーンワードを超えて：中国企業のカーボンウォッシング検出のためのNLP駆動型開示・排出ギャップ指標

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

🔬研究者:Provides a state-of-the-art NLP framework for greenwashing detection with rigorous evaluation, offering a benchmark for future research in climate disclosure analysis.

🏢実務担当者:Offers a tool to assess and improve the credibility of corporate climate disclosures, useful for sustainability teams to benchmark against peers and for investors to inform decisions.

🏛政策担当者:Demonstrates a data-driven approach to monitor corporate carbon washing, which can inform regulatory oversight and enforcement of climate disclosure rules.

Corporate carbon washing — the systematic misalignment between disclosed environmental commitments and verified greenhouse gas emission trajectories — constitutes a critical failure in contemporary climate governance. Existing detection approaches based on manual content analysis and rule-based keyword scoring are insufficient to process the semantic complexity and volume of modern corporate sustainability disclosures. This paper presents CarbonBERT-CN, a contrastively pre-trained, domain-adapted transformer encoder initialized from FinBERT and fine-tuned on an 18,463-entry Chinese climate disclosure lexicon through a three-stage curriculum learning pipeline. CarbonBERT-CN underpins a hierarchical multi-task learning architecture that simultaneously performs (i) green commitment intensity regression, (ii) claim specificity classification (vague / measurable / verified), and (iii) temporal commitment horizon detection (short- / medium- / long-term). The outputs of these three tasks are fused through a gated cross-attention mechanism with company-level node embeddings derived from a heterogeneous knowledge graph linking firms, industry sectors, regulatory policies, and provincial emission monitoring data. The fused representation feeds a scoring module that produces the Disclosure-Emission Gap (DEG) Index — a continuous, interpretable metric quantifying the divergence between a company’s textual environmental posture and its verified emission performance. Evaluated on the China Corporate Climate Disclosure (C3D) dataset — a longitudinal corpus of 3,124 sustainability reports from 1,687 A-share listed companies spanning 2014–2022 — the framework achieves an F1-score of 0.871, AUC-ROC of 0.903, and Macro-F1 of 0.858, with statistically significant F1 improvements of 8.3 to 23.7 percentage points over eight competitive baselines (Bonferroni-corrected McNemar $$p < 0.001$$ across all pairwise comparisons). Ablation studies confirm the individual contributions of curriculum pre-training, H-MTL with PCGrad gradient surgery, and HKG-augmented gated cross-attention fusion. External validity is established through three independent criteria: a within-sector Pearson correlation of 0.753 between DEG and audited Scope 1+2 emission intensity, a Spearman correlation of 0.683 between DEG and inter-agency ESG rating divergence, and a $$+0.327$$ mean DEG difference between firms that did and did not receive subsequent environmental administrative penalties (Mann–Whitney $$p < 0.001$$ ). SHAP-based feature attribution further reveals that text, knowledge-graph, and emission-anchor components contribute 41.2%, 21.8%, and 13.4% respectively to DEG variance, with the remaining 23.6% attributable to the H-MTL task heads. Cross-industry and cross-ownership analyses — robust to industry/province fixed effects, propensity-score matching, and instrumental-variable identification — reveal systematic structural determinants of carbon washing propensity in China’s corporate landscape. Code, the C3D dataset, and the Chinese Climate Disclosure Lexicon are released for peer review at an anonymised repository.

• openalex https://doi.org/10.1007/s44443-026-01075-wfirst seen 2026-08-27 05:08:52