Assessing Corporate Sustainability Alignment through NLP: From Narrative Disclosure to Actionable ESG Scores

Corporate sustainability reports are a rich but highly unstructured source of textual data, making systematic com parison across firms difficult. This paper proposes a text-as-data framework that transforms narrative sustainability disclosures into structured representations aligned with the European Sustainability Reporting Standards (ESRS). Using a Large Language Model in a Chain-of-Verification (CoVe) configuration, we extract a validated corpus of statement-level commitments (promises, goals, and targets), each annotated with topic, quantification, and tempo ral metadata. From this corpus, we construct three analytical layers: (i) a verified commitment inventory, (ii) an alignment profile capturing topic coverage and evidence consistency, and (iii) a sparse KPI presence matrix derived from quantified statements. These layers are aggregated into four interpretable indices—EWDI, DDS, CCI, and TAS—designed to capture complementary latent dimensions of textual disclosure: breadth, measurement depth, credibility, and temporal ambition. All indices are computed through transparent and reproducible transformations of the extracted data, without reliance on additional model-based scoring. The methodology is applied to a corpus of sustainability reports from firms within a supply chain context, en abling multivariate comparison and ranking. Results show systematic decoupling between narrative ambition and quantitative evidence, as well as substantial heterogeneity in disclosure structures across sectors. In particular, the framework identifies a recurring pattern characterized by high commitment density combined with low measurable substantiation, providing a structural signal that can be interpreted as potential greenwashing risk. By formalizing the transformation from unstructured text to structured statistical indicators, this work contributes to the statistical analysis of textual data by bridging information extraction and interpretable aggregation in a reproducible pipeline.