This thesis aims to explore different language models suitable for extracting information from documents, focusing on those trained in the Czech language. The goal is to use and/or adapt English language models to extract data from Czech documents and optimize this process by fine-tuning the models. Since the research is performed in collaboration with the Police of the CR as an application partner, the research will focus on use cases applicable to police investigations.
The research will extract data from documents using the Retrieval Augmented Generation method while preprocessing input documents to remove unnecessary data. It will also aim to deploy and manage models within the infrastructure of Masaryk University, including Kubernetes, and generate training data from provided documents.
The thesis will evaluate the results obtained from the research and develop a robust evaluation methodology. It will also tentatively explore the possibilities for fine-tuning the available models on the Czech language, at least partially, or performing the use-case research challenges just on analogous English documents.
The thesis will provide a comprehensive overview of the field's current state, contribute to understanding and applying language models for information extraction from Czech documents, evaluate the results rigorously, and discuss potential future directions for this line of research.