Automatické vyhodnocování výročních zpráv firem z hlediska zelených úvěrů

This thesis focuses on the automated evaluation of corporate reports for the purposes of green loan assessment. In the context of the growing importance of sustainable finance and EU regulatory requirements, banks increasingly need to efficiently obtain ESG indicators, which are often available only in annual reports and sustainability reports. The aim of this thesis is to design and evaluate an approach for the automated extraction of selected EU Taxonomy indicators from these documents. Two approaches are analyzed: a traditional rule-based approach and a modern approach based on large language models (LLMs). The proposed solution includes the design of a methodology, its implementation in Python, and evaluation on a dataset of real corporate reports. The results show that the rule-based approach faces significant limitations due to the variability of documents and the complexity of the PDF format. In contrast, the LLM-based approach achieves high extraction accuracy. The best-performing configuration correctly extracted 98% of the monitored indicators, and for 65% of the documents all required indicators were extracted without errors. At the same time, a significant reduction in time and cost requirements compared to manual processing was achieved. The thesis confirms that the use of LLMs represents an effective tool for automating document processing in the financial sector and can contribute to more efficient processes related to sustainability assessment and green financing. The thesis also includes a Python implementation that enables replication and further development of the proposed approach.