Integrating Large Language Models and Knowledge Graphs for Financial Applications: Challenges, Solutions, and Future Directions

The presentation begins with an overview of the growing importance of large language models (LLMs) and knowledge graphs (KG) in the financial domain, highlighting the high standards for accuracy, explainability, and regulatory compliance that financial AI must meet.

It identifies key challenges: traditional Retrieval‑Augmented Generation (RAG) struggles with domain‑specific queries, LLMs suffer from hallucinations, and building comprehensive KGs incurs high cost and complexity, especially in relation extraction.

To address these issues, several solution pathways are proposed. A soft‑coupling approach uses the KG as an external knowledge source for the LLM, feeding retrieved triples as context. A tight‑coupling approach treats the LLM as an intelligent agent that iteratively explores the KG, performing multi‑hop reasoning. The "Think on Graph" paradigm combines both ideas, enabling dynamic graph traversal guided by user intent.

The talk introduces the Context Graph (or "语境图谱") technique, which avoids explicit relation extraction by linking documents to entity co‑occurrence edges, dramatically reducing construction cost and improving scalability. Comparisons with GraphRAG show orders‑of‑magnitude speed and resource advantages, as Context Graph relies on lightweight BERT‑based entity extraction and vector similarity rather than heavy LLM‑driven summarisation.

Multimodal extensions are discussed, showing how tables, images, and charts can be embedded and attached to entity edges, enabling a unified multimodal KG. Future work envisions multimodal graph integration, agentic financial systems that combine LLM reasoning with graph‑guided planning, and decision‑making applications across investment, risk management, and product design.

Practical applications are illustrated through the "Economic Superbrain" platform, the Alpha‑GPT factor‑mining system, and collaborations with digital‑human providers, demonstrating real‑world impact and competitive results in contests such as WorldQuant and ICLR.

The session concludes with a Q&A covering graph construction, search strategies, multimodal handling, and the roadmap toward more explainable, efficient, and agentic AI for finance.