Powered SQL education: Automating SQL/PLSQL question classification with LLMs and machine learning

Abstract

Mastering Structured Query Language/Procedural Language (SQL/PLSQL) is considered challenging for academic students and industrial professionals, showing a significant gap between academic preparation and industrial demands that leads both to seek solutions on Stack Overflow (SO). This research presents a novel automated framework to classify SQL/PLSQL questions and shed light on learning challenges. A new dataset was collected from SO posts, totaling 10,266 questions, and categorized into five categories—Data Definition Language (DDL), Data Manipulation Language (DML), Data Query Language (DQL), Data Control Language (DCL), and Transaction Control Language (TCL)—using the LLM GPT-4o-mini API, followed by preprocessing and applying Machine Learning (ML) techniques like Random Forest and XGBoost. Results show that Data Query Language (DQL) and Data Manipulation Language (DML) are the most challenging areas, with Random Forest and XGBoost producing the highest classification accuracy at 85.57% and 85.13%, respectively, while DDL and DCL appear less often. This research bridges the gap between academic and industrial requirements, concluding that AI-driven analysis identifies the real challenges, suggesting that the academic curriculum enhance hands-on problem-solving to meet industry needs.