Integrating Large Language Models and Knowledge Graphs for Intelligent Financial Regulatory Risk Identification

Abstract

This paper addresses the complexity and uncertainty in financial regulatory risk identification and proposes an intelligent recognition mechanism that integrates large language models with knowledge graphs. The study first analyzes the limitations of traditional methods in handling multi-source heterogeneous financial data. It points out that relying only on semantic modeling or rule-based constraints cannot fully capture potential risk signals. To overcome this, the proposed framework combines the semantic understanding of large language models with the structured reasoning of knowledge graphs. It is designed to model both unstructured text information and entity relationship networks simultaneously. In terms of method design, a self-attention mechanism is introduced to enhance contextual modeling. A graph convolutional network is used to embed and propagate entities and relations within the knowledge graph. A fusion strategy is then applied to achieve complementarity between semantics and knowledge. This ensures both accuracy and interpretability in risk signal identification. In the experimental part, the study conducts multidimensional comparisons and sensitivity analyses, including hyperparameter settings, environmental conditions, and data perturbations, to validate the effectiveness and robustness of the proposed method. The results show that the method outperforms traditional models in core metrics such as AUC, ACC, F1-Score, and Precision. It also maintains stable performance under various conditions, demonstrating its ability to adapt effectively to complex financial regulatory scenarios. This study not only confirms the rationality of semantic and knowledge integration but also highlights its potential in processing multi-source financial information and improving identification accuracy. It provides a new technical path for building intelligent frameworks for risk identification.