From Text Analysis to Metrics: AI Approaches to ESG Assessment

Abstract

In ESG assessments, disclosure-centric scorecards are being replaced by AI-based sentence-level systems, which offer organized aggregate based on precise criteria. The shift from topic models and human supervision to domain-adaptive Transformer classifiers and evidence-based language models with retrieval capabilities is traced in this work. In order to create trustworthy textual metrics from news, earnings calls, and reports, we concentrate on multi-label phrase annotation by combining non-ESG controls, industry materiality, and financial sentiment. The evaluation takes precision/recall, external validity, and structural variations among rating agencies into account. Decision-making applications in supply chain visibility, investment research, and disclosure assurance are covered, as well as governance elements like documentation, energy use, and human-computer interaction. We conclude by outlining potential avenues for future research and development, such as cross-lingual robustness, retrieval stability, benchmark design, and low-carbon AI. Our suggested "text → metrics → econometrics" approach complements current ratings while improving the scalability, timeliness, comparability, and auditability of ESG measurement.