LLM-Assisted ESG Data Extraction from Corporate Reports

This thesis focuses on the development and assessment of a comprehensive pipeline aimed at extracting Environmental, Social, and Governance (ESG) key performance indicators (KPIs) from corporate websites and sustainability reports. The proposed pipeline integrates web discovery, a Large Language Model (LLM) for information extraction, and a deterministic post-processing mechanism to standardize noisy outputs into structured, analysis-ready tables. A case study carried out at Illuminem is used to evaluate field-level accuracy, with an exploration of challenges stemming from unstructured data formats, variability in model outputs, and inconsistencies in reporting practices.