Improving Cypher Query Generation from Natural Language using Reinforcement Learning with Large Language Models

Abstract

This study addresses the automated generation of Cypher queries from natural language expressions for graph databases, systematically analyzing the impact of various training strategies and data augmentation techniques on model performance. Although Large Language Models (LLMs) demonstrate significant potential in natural language-to-query tasks, their generalization to unseen schemas and structural accuracy remain critical challenges. In this context, the study investigates the contributions of Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and cross-domain training strategies utilizing relational database (SQL) data to the Text-to-Cypher task.