Dynamic Taxonomy Construction and Thematic Filtering for Financial Knowledge Graphs

This paper presents a framework that leverages Large Language Models (LLMs) to dynamically construct financial taxonomies and knowledge graphs from unstructured sources, specifically daily conference call transcripts. Our approach addresses the high variability in how financial entities and relationships are described by mapping extracted node types and relations to their parent nodes. A key innovation is generating an adaptive taxonomy for theme generation and applying thematic filtering, which organizes information by relevant topics and enhances the efficiency of data analysis and search for investors. By classifying and filtering graph data according to themes, our framework reduces complexity and variability, making financial information more accessible and actionable. Empirical results demonstrate improvements in graph search efficiency and enable the extraction of richer insights from financial data, supporting more effective decision-making in dynamic market environments.