NormaScore: Intelligent Database Normalization Assessment Framework for DBMS Education

Normalization is a core concept in Relational Database Management Systems (RDBMS) education, but assessing student’s normalization solutions is tedious and hard to scale. Current automated methods produce schemas out of structured inputs which cannot be used to do a mathematically precise and staged evaluation. This work introduces an end-to-end intelligent framework NormaScore that automatically evaluates database normalization solutions from First Normal Form (1NF) to Third Normal Form (3NF). The framework integrates a domain-specific Qwen2.5-7B-Instruct model trained with Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) techniques, as well as deterministic algorithms for computing attribute closure, verifying joins without losing all data elements, and preserving functional dependencies. NormaBench is a first formally validated dataset of 113 problems from various real-world domains for normalization reasoning, which is a significant contribution. Reinforcement learning is guided by a mathematically sound reward function and correct normalization reasoning is enforced by stage-specific rewards. Experimental results show that the reward is improved by 0.26–0.53 compared with the REINFORCE baseline, and BERTScore F1 is 0.87–0.92 for reference schema generation, while the Pearson correlation is 0.91 with expert assessment of students. The results illustrate the effectiveness of the proposed solution, NormaScore, in automated assessment of normalization degree in databases, with respect to accuracy and scalability.