Template-based Explainable Inference over High-Stakes Financial Knowledge Graphs

Abstract

Declarative query languages based on logic programming, like Datalog and its extensions, have recently found successful applications in modeling complex knowledge-based scenarios, such as reasoning over Enterprise Knowledge Graphs (EKG), by encoding business rules to derive new valuable knowledge. Presenting this derived knowledge with comprehensible natural language explanations is paramount to increasing transparency, accountability, and fairness in AI-based systems. While Large Language Models (LLMs) offer promising directions, full industrial adoption in critical settings requires a trustworthy solution that ensures both accurate, clear explanations and compliance with strict data protection standards (i.e., by not sharing data with third parties). This work introduces a novel approach for the generation of textual explanations from data-driven inference processes where data protection is crucial, such as in sensitive financial applications governed by deductive rules encoded by the Central Bank of Italy. We propose a static structural analysis method that identifies a finite set of reasoning patterns from business rules, which are then used to generate fluent natural language explanations. By capturing the main interconnections between rules, our approach generates explanations comparable in quality to those produced by LLMs, but without requiring data sharing through external APIs or cloud servers, thus ensuring data protection in high-stakes, sensitive applications. Furthermore, our method guarantees that explanations are both correct and complete, unlike LLM-generated ones, which may suffer from critical omissions.