Post-Hoc Natural-Language Explanations of Component-Based Knowledge Graph Question Answering Systems Generated by LLMS

This is the authors' abstract. We don't add key points for this paper.

Modern software systems have become so complex that explaining their decisions poses significant challenges for both developers and users.This article addresses the explainability of component-based Knowledge Graph Question Answering (KGQA) systems, where components often rely on AI-driven processes.Such processes can be opaque, making it difficult even for KGQA experts to interpret the underlying behavior and outcomes.To tackle this issue, we propose an approach that leverages the input and output data flows of system components as a basis for representing their behavior and generating explanations.This enables users to better understand how decisions are made.In the KGQA framework considered here, component data flows are expressed as SPARQL queries (inputs) and RDF triples (outputs).Consequently, our work also provides insights into verbalizing these data types.Through experiments, we evaluate our approach, comparing template-based explanation generation (baseline) with automatic generation using Large Language Models (LLMs) configured in various ways.The results demonstrate that LLM-generated explanations are of high quality and generally outperform template-based methods based on user evaluations.This approach thus facilitates the automated natural-language explanation of KGQA components' behavior and decisions, contextualized within RDF and SPARQL representations.