Dynamic taxonomy Construction and Thematic Filtering for Financial Knowledge Graphs with LLMs

Atousa Assadihaghi, Reza Amini, Daniel Schiemert, Ewan Mackie, Zhiwen Xu

Keywords: Taxonomy, Theme, Knowledge graph, LLM

TL;DR: This paper presents a novel framework that leverages large language models to dynamically construct ontologies and apply thematic filtering, significantly enhancing the consistency and analytical value of financial knowledge graphs.

Abstract: Financial analysis and investment strategies increasingly rely on integrating diverse data sources, such as text, numerical data, and domain-specific knowledge. However, the variability in how financial entities and relationships are described poses significant challenges for data consistency and analysis. This paper introduces a novel approach that uses Large Language Models (LLMs) to dynamically construct taxonomies and financial knowledge graphs from sources like daily conference call transcripts. By mapping extracted entities and relationships to parent nodes within adaptive taxonomy, our method reduces data variability and enhances graph consistency. A key innovation is the introduction of theme-based and thematic filtering, which organizes information by relevant topics, streamlining data analysis and improving search efficiency for investors. Experimental results demonstrate that this approach not only reduces complexity and variability but also improves graph search efficiency by 6% while extracting richer insights from financial data.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 16530