System and Method of Natural Language to SQL Translation Using Neural Network Architecture

Abstract

The translation of natural language into SQL (NL2SQL) plays a vital role in making database systems accessible to non-technical users. Although large language models (LLMs) have significantly advanced NL2SQL performance, deploying them on hardware with limited computational resources remains a major challenge. This paper presents an optimized approach based on the Gemma3 1B model, designed specifically for efficient NL2SQL translation under consumer-grade hardware constraints. The proposed method integrates parameter-efficient fine-tuning (PEFT) using QLoRA, schema-aware processing, and prompt engineering to enhance both accuracy and performance. Experimental evaluations conducted on benchmark datasets such as Spider and WikiSQL demonstrate notable improvements in query translation accuracy despite hardware limitations. This study provides a practical framework for implementing lightweight yet effective NL2SQL systems, thereby promoting wider accessibility and democratization of database interaction.