cs.CLMar 29, 2026

ESGLens: An LLM-Based RAG Framework for Interactive ESG Report Analysis and Score Prediction

Authors: Tsung-Yu Yang, Meng-Chi Chen

Organizations: Massachusetts Institute of Technology

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Environmental, Social, and Governance (ESG) reports are central to investment decision-making, yet their length, heterogeneous content, and lack of standardized structure make manual analysis costly and inconsistent. We present ESGLens, a proof-of-concept framework combining retrieval-augmented generation (RAG) with prompt-engineered extraction to automate three tasks: (1)~structured information extraction guided by Global Reporting Initiative (GRI) standards, (2)interactive question-answering with source traceability, and (3)ESG score prediction via regression on LLM-generated embeddings. ESGLens is purpose-built for the domain: a report-processing module segments heterogeneous PDF content into typed chunks (text, tables, charts); a GRI-guided extraction module retrieves and synthesizes information aligned with specific standards; and a scoring module embeds extracted summaries and feeds them to a regression model trained against London Stock Exchange Group (LSEG) reference scores. We evaluate the framework on approximately 300 reports from companies in the QQQ, S&P500, and Russell1000 indices (fiscal year 2022). Among three embedding methods (ChatGPT, BERT, RoBERTa) and two regressors (Neural Network, LightGBM), ChatGPT embeddings with a Neural Network achieve a Pearson correlation of 0.48 (R2≈0.23R^{2} \approx 0.23) against LSEG ground-truth scores -- a modest but statistically meaningful signal given the ∼300{\sim}300-report training set and restriction to the environmental pillar. A traceability audit shows that 8 of 10 extracted claims verify against the source document, with two failures attributable to few-shot example leakage. We discuss limitations including dataset size and restriction to environmental indicators, and release the code to support reproducibility.

Figures & tables

Figure 1: Comparison between General AI-powered PDF tools and the proposed Interactive Question-Answering System for ESG Reports. (a) General AI-powered PDF tools typically take user-uploaded PDF files, process them with an unknown embedding model (primarily focused on text with limited support for equations, tables, charts, and images), and submit them to a commercial LLM using a generic prompt along with the user’s questions, producing answers with unclear processing steps. (b) Our framework begins with the user selecting specific companies from a database to compare their ESG reports. The user then submits questions about one or more companies’ ESG data. A crafted prompt tailored specifically to ESG topics is generated and sent to a commercial LLM (ChatGPT) via an API, minimizing post-processing. The resulting answer is delivered directly to the user interface.
Figure 2: Detailed process framework of ESGLens, illustrating the five-stage pipeline. (1) Data Collection : selecting companies from the QQQ, S&P 500, and Russell 1000 indices and downloading their ESG reports alongside LSEG reference scores; (2) PDF Processing : splitting reports into chunks via RecursiveCharacterTextSplitter, embedding them with OpenAI embeddings, and storing them in a FAISS vector database; (3) Data Extraction : -querying the vector database with GRI standards to retrieve the top 20 relevant chunks per standard; (4) Summarization : combining GRI standards, retrieved context, and prompt guidelines as input to ChatGPT to produce structured summaries; and (5) Report Scoring : embedding the summaries and feeding them into a regression model (Neural Network or LightGBM) to predict a quantitative ESG score, which is then benchmarked against LSEG ground-truth scores.
PromptStructureResult
1Instruction, question, context, and a final instruction.Fair. Only one result was generated.
2Instruction, context, and question.Good. More comprehensive results, but lacked a specified output format.
3Instruction with examples, context, and question.Best. Generated a clean and structured summary.
Table 1: Comparison of prompt structure and results.
Figure 3: (a) Training loss of NN. (b) Correlation of predicted and actual ESG scores using ChatGPT, BERT, and RoBERTa embeddings with Neural Network and LightGBM models. ChatGPT outperforms the others.
ItemExtracted InformationParagraph in DocumentCorrectness
1The annual total energy reduction reached 280 GWh, and the cumulative total reached 810 GWh.… obtained an average grade of B-, exceeding the initial target of C. TSMC requires and assists suppliers to improve their sustainability performance. In 2023, the annual total energy reduction reached 280 GWh, and the cumulative total reached 810 GWh. The annual total water reduction reached 13.5 million metric tons, and the cumulative total reached 42.58 million metric tons. Additionally, the waste production among major …✓
2The company implemented 822 energy-conservation programs, resulting in a total reduction of 800 GWh in electricity consumption.… Made significant strides in climate change mitigation, achieving net-zero emissions in scope 1 and scope 2 at TSMC overseas operations; expanded its onshore wind farm power supply, increasing the proportion of renewable energy usage from 40% to 60% by 2030, and proposed 822 energy-conservation programs, resulting in a total reduction of 800 GWh in electricity consumption. …✓
3The company reported a 5% increase in renewable energy use, reaching 20% of total energy consumption.(No related information in the ESG report.)✗
4The total energy consumption for electricity alone was 500 GW.(No related information in the ESG report.)✗
5The company planned and carried out 822 energy-saving measures across eight major categories, saving an additional 830 GWh.… Planned and carried out 822 energy-saving measures across eight major categories, saving an additional 830 GWh. See Increase Energy Efficiency for more information.✓
6The company missed the annual target for the energy efficiency of 5nm volume production.… In 2023, TSMC consumed a total of 24,700 GWh in energy, which resulted in a 0.1% shortfall from its 5nm production efficiency target. This was mainly due to global supply chain disruptions. The failure to achieve this target prompted TSMC to invest further in energy-saving measures and implement additional guidelines.✓
Table 2: Comparison of extracted information and documented statements with correctness verification.
CategoryExtracted InformationToneValue
emissions_1Scope 1 (gross emissions) for fiscal year 2021 was 55,200 metric tons CO2e.Neutral55,200
emissions_1Scope 2 (market-based) emissions for fiscal year 2021 were 2,780 metric tons CO2e.Neutral2,780
emissions_1Scope 3 (gross emissions) for fiscal year 2021 were 23,130,000 metric tons CO2e.Neutral23,130,000
emissions_2The company targets a 30% reduction in carbon emissions by 2025.Positive30%
emissions_3Manufacturing emissions account for approximately 70% of the company’s gross carbon footprint.Positive70%
emissions_4Manufacturing emissions account for approximately 70% of the company’s gross carbon footprint.Negative—
Table 3: Categorized summary of extracted ESG information with tone and value assessment.
CategorySummaryToneValuePage
emissions_1Scope 1 (gross emissions) for fiscal year 2021 was 55,200 metric tons CO2e.Neutral55,20078
emissions_1Scope 2 (market-based) emissions for fiscal year 2021 were 2,780 metric tons CO2e.Neutral2,78078
emissions_1Scope 3 (gross emissions) for fiscal year 2021 were 23,130,000 metric tons CO2e.Neutral23,130,00078
emissions_4Manufacturing emissions account for approximately 70% of the company’s gross carbon footprint.Negative—14
emissions_4Several components require manufacturing processes using fluorinated gases with high global warming potential.Negative—16
resource_1Total energy consumption and proportion from renewable sources (2021): 18,100,000 MWh (100% renewable); (2020): 11,400,000 MWh (100% renewable); (2019): 5,700,000 MWh (100% renewable); (2018): 4,100,000 MWh (99% renewable); (2017): 1,900,000 MWh (97% renewable).Neutral—16
Table 4: Environmental Data Summary
CategorySummaryToneValuePage
emissions_1Compared to the previous year, overall Scope 3 emissions decreased by approximately 25%, driven by reduced production volume in 2022, increased renewable energy usage, and updated EPA emissions factors.Positive—25
emissions_2The company targets a 55% reduction in Scope 3 emissions per million USD of value added by 2030.Positive—10
emissions_3The company aims to reduce emissions by approximately 50% by 2030 and achieve net-zero emissions thereafter.Positive—10
emissions_4The company targets a 30% reduction in carbon emissions by 2025.Positive30%3
resource_1The company aims to increase the share of onsite generation and externally supplied clean electricity.Positive—19
resource_2The company targets a 30% reduction in carbon emissions by 2025.Positive30%3
Table 5: Summary of emissions goals and reported achievements.
CategoryQuestions
emissions_1What are the greenhouse gas emissions for Scope 1, Scope 2, and Scope 3?
emissions_2What is the carbon intensity (emissions per unit of product or revenue)?
emissions_3What are the carbon reduction targets, and what progress has been made?
emissions_4How are air pollutants, such as nitrogen oxides and sulfur oxides, managed?
resource_1What is the total energy consumption, and what proportion derives from renewable sources?
resource_2How is water used, including total consumption and recycling rate?
Table 6: Environmental Questions by Category
More in this topic

Document Information Extraction

Document Information Extraction turns text, tables, forms, and images into structured, searchable data.

See all 290 papers

Explore similar work

CardsList
Sep 14, 2026

Empirical Evaluation of Open-Source Large Language Models for Retrieval-Augmented Generation in ESG Domain

Environmental, Social, and Governance (ESG) reporting is critical for corporate accountability, with Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) offering strong potential to automate KPI extraction. However, open-source LLM performance in domain-specific ESG tasks remains insufficiently understood. This paper evaluates open-source LLMs in ESG contexts using a structured framework and evaluation resource based on 498 real-world ESG reports from EU-listed companies (2010-2024). We evaluate seven open-source models (2B to 30B parameters) -- glm-4.7-flash, nemotron-3-nano:4b, qwen3:4b-instruct, gemma3:4b, gemma4:e4b, gemma4:e2b, and ministral-3:8b -- using 100 persona-based synthetic QA pairs covering ESG information needs. System performance is assessed via RAGAS metrics, including contextual recall, precision, relevance, faithfulness, answer relevancy, and factual correctness. Results show notable performance variations across architectures. Retrieval performance is strong across models (context recall around 0.58-0.61, context precision around 0.78-0.81, context relevance 0.965-0.985). Generation diverges most on faithfulness (0.607-0.822) and least on answer relevancy (0.760-0.881): glm-4.7-flash leads in faithfulness (0.822), qwen3 in factual correctness (0.449), and ministral-3 in answer relevancy (0.881). Low overall factual correctness (0.387-0.449) highlights the need for domain-specific fine-tuning. This work provides data-driven guidance for deploying open-source models in ESG reporting.

Motaz Saad, Anna Borrelli, Ivan Gentile +3

University of Salento · University of Naples Federico II · IFAB Foundation

LLM EvaluationLLM GroundingLLM Information ExtractionRAG Evaluation
AskPDF
Mar 13, 2026

ESG-Bench: Benchmarking Long-Context ESG Reports for Hallucination Mitigation

As corporate responsibility increasingly incorporates environmental, social, and governance (ESG) criteria, ESG reporting is becoming a legal requirement in many regions and a key channel for documenting sustainability practices and assessing firms' long-term and ethical performance. However, the length and complexity of ESG disclosures make them difficult to interpret and automate the analysis reliably. To support scalable and trustworthy analysis, this paper introduces ESG-Bench, a benchmark dataset for ESG report understanding and hallucination mitigation in large language models (LLMs). ESG-Bench contains human-annotated question-answer (QA) pairs grounded in real-world ESG report contexts, with fine-grained labels indicating whether model outputs are factually supported or hallucinated. Framing ESG report analysis as a QA task with verifiability constraints enables systematic evaluation of LLMs' ability to extract and reason over ESG content and provides a new use case: mitigating hallucinations in socially sensitive, compliance-critical settings. We design task-specific Chain-of-Thought (CoT) prompting strategies and fine-tune multiple state-of-the-art LLMs on ESG-Bench using CoT-annotated rationales. Our experiments show that these CoT-based methods substantially outperform standard prompting and direct fine-tuning in reducing hallucinations, and that the gains transfer to existing QA benchmarks beyond the ESG domain.

Siqi Sun, Ben Peng Wu, Mali Jin +3

School of Computer Science, University of Sheffield, Sheffield, UK

LLM Hallucination MitigationLLM Safety BenchmarksDocument QALong-Context QA
AskPDF
Mar 31, 2026

Towards Empowering Consumers through Sentence-level Readability Scoring in German ESG Reports

With the ever-growing urgency of sustainability in the economy and society, and the massive stream of information that comes with it, consumers need reliable access to that information. To address this need, companies began publishing so called Environmental, Social, and Governance (ESG) reports, both voluntarily and forced by law. To serve the public, these reports must be addressed not only to financial experts but also to non-expert audiences. But are they written clearly enough? In this work, we extend an existing sentence-level dataset of German ESG reports with crowdsourced readability annotations. We find that, in general, native speakers perceive sentences in ESG reports as easy to read, but also that readability is subjective. We apply various readability scoring methods and evaluate them regarding their prediction error and correlation with human rankings. Our analysis shows that, while LLM prompting has potential for distinguishing clear from hard-to-read sentences, a small finetuned transformer predicts human readability with the lowest error. Averaging predictions of multiple models can slightly improve the performance at the cost of slower inference.

Benjamin Josef Schüßler, Jakob Prange

University of Augsburg, Germany · German Center for Addiction Research in Childhood and Adolescence (DZSKJ), University Medical Center Hamburg-Eppendorf, Germany

Fine-Tuning
AskPDF