ESG company proiling, a Retrieval-Augmented Generation (RAG): Approach for data-driven corporate analysis

This Master’s thesis addresses the challenge of creating comprehensive fact-based ESG (Environmental, Social, and Governance) company profiles, a procedure obstructed by the manual synthesis of diverse, often unstructured, data, and the limitations of standard Large Language Models (LLMs) in ensurin...

Descripción completa

Detalles Bibliográficos
Autor: Aguirre Calderón, Juan
Tipo de recurso: tesis de maestría
Fecha de publicación:2025
País:España
Institución:Consejo Superior de Investigaciones Científicas (CSIC)
Repositorio:O2, repositorio institucional de la UOC
OAI Identifier:oai:openaccess.uoc.edu:10609/154180
Acceso en línea:https://hdl.handle.net/10609/154180
Access Level:acceso abierto
Palabra clave:ESG (Environmental, Social, Governance)
company profile
Retrieval-Augmented Generation (RAG)
Large Language Models (LLMs)
perfil ESG (Ambiental, Social y Gobernanza) de empresa
Generación Au-mentada de Recuperación (RAG)
Grandes Modelos de Lenguaje (LLMs)
Natural language processing (Computer science) -- FMDP
Tractament del llenguatge natural (Informàtica) -- TFM
Descripción
Sumario:This Master’s thesis addresses the challenge of creating comprehensive fact-based ESG (Environmental, Social, and Governance) company profiles, a procedure obstructed by the manual synthesis of diverse, often unstructured, data, and the limitations of standard Large Language Models (LLMs) in ensuring accuracy and traceability. The project proposes and implements a Retrieval-Augmented Generation (RAG) framework designed to automate the initial drafting of ESG profiles while maintaining high factual integrity and source attribution. The implemented methodology involves the ingestion of external documents (such as sustainability and research reports) and proprietary internal ESG metrics. External documents undergo pre-processing, including robust parsing and chunking, before being vectorized and stored in a knowledge base. The core RAG pipeline dynamically retrieves relevant document pages based on the specific ESG topic being analyzed, incorporates formatted internal data, and uses sophisticated prompt engineering to guide a generative LLM. A crucial post-generation verification layer uses a separate LLM to automatically assess factual accuracy and citation correctness against sources. Evaluation of the implemented system has shown significant efficiency gains, reducing the time required for analysts to produce a final ESG company profile by approximately 80% compared to traditional manual methods. The automated verification layer has successfully identified inconsistencies and errors, highlighting its value for output reliability. Analysts feedback has confirmed the utility of the drafts and has praised their integrated traceability features (citations, access to sources via the user interface) to facilitate the validation. This architecture is applicable for evidence-based synthesis from heterogeneous data sources.