G-SQL: A Schema-Aware and Rule-Guided Approach for Robust Natural Language to SQL Translation

Abstract

This paper introduces G-SQL, a schema-aware and rule-guided framework for translating Natural Language Queries (NLQ) into SQL, designed to support users with limited technical expertise. G-SQL bridges the limitations of traditional rule-based and deep learning approaches by combining JSON-based schema serialization with Natural Language Processing (NLP) tools, such as GloVe and Stanford CoreNLP. The system encodes database schemas—including tables, columns, keys, and relationships—into a structured JSON format, enabling efficient and interpretable mapping between query tokens and schema elements. SQL statements are generated using a guided template-driven strategy. Evaluated across three benchmark datasets (IMDB, Yelp, MAS), For easy queries, it achieved 100% execution accuracy on all three datasets. On medium queries, G-SQL reached 95% on IMDB, 85% on Yelp, and 90% on MAS. For hard queries, the accuracy was 78% on IMDB, 65% on Yelp, and 72% on MAS. Finally, on extra-hard queries, G-SQL achieved 55% on IMDB, 45% on Yelp, and 50% on MAS., demonstrating robust performance across varying query complexities.