Discriminative meets generative: Automated information retrieval from unstructured corporate documents via (large) language models

Abstract

This paper demonstrates the potential of (large) language models to transform accounting practice and research by automating information retrieval from unstructured sources. While information retrieval in accounting has hitherto been predominantly addressed through handcrafted, rule-based systems, we have devised an approach based entirely on machine learning methods from the field of natural language processing (NLP). Specifically, we consider and contrast two modeling paradigms in NLP: discriminative modeling and generative modeling. In the former case, we fine-tune a language model pre-trained for multilingual, visually rich document understanding. In the latter, we apply prompting techniques to utilize a large language model (LLM) without prior training. We illustrate our approach by retrieving group ownership data from annual reports in Portable Document Format (PDF). We successfully retrieve group ownership information for both modeling paradigms, achieving a strong overall accuracy with a low percentage of false negatives. Furthermore, we consider the impact of different reporting languages and accounting standards.