Effectiveness of Explainable AI for interpreting NL2SQL systems

Ascertain the effectiveness of applying explainable AI techniques—such as surrogate models and saliency maps—to interpret and validate the decision-making processes of NL2SQL systems, particularly when combining Large Language Models (LLMs) and Pre‑trained Language Models (PLMs).

Background

To build trustworthy NL2SQL solutions, the authors propose using explainable AI (XAI) methods to understand why models generate specific SQL, thereby increasing transparency and reliability.

The paper explicitly states that whether such XAI techniques are effective in the NL2SQL setting—especially when LLMs and PLMs are combined—remains unknown, defining a clear unresolved question about evaluation methodology and practical utility.