📄 CLARIESG: An End-to-End System for ESG Analysis over Complex Tables in Corporate Reports

Summary

Corporate sustainability reports are a major source of ESG data, but their tables are irregular, with merged headers, nested indicators and company-specific schemas. Their indicators also use inconsistent units, and many questions require arithmetic across several tables or reports. Existing LLM-based tools rarely address all of these issues together. The authors present CLARIESG, a demo system with two layers. A data preparation layer finds the company and its sector through Wikidata, locates GRI-relevant pages with hybrid retrieval, and extracts tables using an OCR ensemble. An LLM then filters and repairs those tables, and users can correct the results through a human-in-the-loop interface. An analytics layer answers questions with NormPoT, a prompting scheme that filters candidate tables, derives unit-conversion rules and then writes and runs Python code. On the multi-table GRI-QA benchmark, NormPoT clearly beats Chain-of-Thought, and CLARIESG with gpt-5-mini far outperforms ChatGPT 5.1 given the full reports. This makes ESG benchmarking and greenwashing claim checks more accurate and auditable.

Key points

• Data preparation pipeline:
  • Retrieval combines multilingual-e5-large-instruct embeddings with TF-IDF.
  • Tables are extracted by an ensemble of Unstructured and Tesseract.
  • An LLM refines tables in two steps: a GRI relevance check, then structure and format correction.
  • The interface is built with Gradio 5.46.0 and calls the OpenAI API.
• Demo scenarios: analysts can ask natural-language questions and receive answers with scorecards, or submit claims for verification against evidence in the reports.
• One-table GRI-QA results (normalized EM):
  • Plain CoT scores best on average (72.7).
  • PoT drops 7.0 points and NormPoT drops 7.9.
  • On the quant subset, however, PoT and NormPoT gain about 13 points.
• Multi-table GRI-QA results: NormPoT reaches 48.4 average EM, 16.0 points above CoT (32.4). PoT gains 8.8 points.
• Comparison with ChatGPT 5.1 (42.2 EM, first 50 questions per multi-table subset):
  • CLARIESG with gpt-4o-mini scores 49.1, but falls to 35.8 when irrelevant "noisy" tables are added.
  • With gpt-5-mini it reaches 79.1 on clean tables and 77.6 with noisy tables.