Transforming Unstructured Sensitive Information into Structured Knowledge

Information is crucial in today's context, yet less than 20% of companies utilize their unstructured data due to its complexity. Information Extraction (IE) is vital for effective data use, but current IE models face four major issues. First, they often provide limited information, such as a simple entity-attribute relation. Second, they struggle with multiple languages. Models like GPT, Mistral, and Llama3 show promise but face a third issue: output reliability due to hallucinations. Fourth, there is a challenge in reducing sensitive data leakage after fine-tuning models. This study introduces an enhanced approach for fine-tuning GPT-based models, designed to extract and assess information involving multiple entities and attributes, performing both multientity extraction (MEE) and multirelation extraction (MRE), and presenting results in a JSON format. Our methodology evaluates the impact of using synthetic data for fine-tuning to ensure reliable outcomes. Applied to legal documents from the Luxembourg Business Registers (LBR), our findings show that replacing sensitive data with synthetic data significantly improves the fine-tuning of Llama3-based models, though not for Mistral-based models. Our top models outperform Mistral in various scenarios, requiring only 500 samples for fine-tuning and running efficiently on modest servers. This approach is suitable for multilingual Information Extraction in any domain.