Beyond ratings: deconstructing ESG disclosure through AI-driven action-level knowledge graph analysis

Third-party environmental, social and governance (ESG) ratings dominate corporate sustainability evaluation despite low inter-rater reliability (average correlation 0.61) and methodological opacity that obscures underlying evidence and weighting assumptions. This study uses large language models (LLMs) to extract granular action records from corporate disclosures and constructs knowledge graphs encoding relationships between companies, topics, stakeholders and activities. From 2024 sustainability reports of 141 Japanese listed firms, analysis extracted 77,734 action records comprising eight elements (company, topic, actor, action, target, modality, evidence and confidence), with 97.7% exceeding 0.9 confidence.Analysis reveals patterns aggregate ratings obscure. Japanese firms emphasize social disclosure (55.4%) over environmental (26.9%), prioritize ongoing implementation (65.5%) over future commitments (19.1%), and concentrate stakeholder attention on employees rather than external groups. Network analysis identifies topic co-occurrence structures and three distinct disclosure archetypes aligned with industry materiality. This framework complements evaluative ratings by enabling source-traceable evidence review, user-customizable materiality weighting and systematic disclosure pattern analysis – capabilities proprietary rating methodologies do not provide. Methodologically, this study contributes scalable infrastructure for action-level ESG analysis. Theoretically, this approach operationalizes stakeholder salience and signaling credibility through empirical extraction of actor–target relationships and temporal commitment patterns.