NAT-NL2GQL: A Novel Multi-Agent Framework for Translating Natural Language to Graph Query Language

ACL ARR 2025 February Submission3946 Authors

Abstract:

There has been increasing interest in using Large Language Models (LLMs) for translating natural language into graph query language (NL2GQL). While progress has been made, current approaches often fail to fully exploit the potential of LLMs to autonomously plan and collaborate on complex NL2GQL tasks.
To address this gap, we propose NAT-NL2GQL, an innovative multi-agent framework for NL2GQL translation. The framework consists of three complementary agents: the Preprocessor agent, the Generator agent, and the Refiner agent. The Preprocessor agent handles tasks such as entity recognition, query rewriting, and schema extraction. The Generator agent, a fine-tuned LLM trained on NL-GQL data, generates corresponding GQL statements based on queries and their related schemas. The Refiner agent refines the GQL or context using error feedback from the GQL execution results.
In the absence of high-quality open-source NL2GQL datasets based on nGQL syntax, we developed StockGQL, a Chinese dataset derived from a Chinese financial market graph database, which will be made publicly available to support future research.
Experiments on the StockGQL and SpCQL datasets demonstrate that our approach significantly outperforms baseline methods, underscoring its potential to drive advancements in NL2GQL research.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: Multi-Agent, Large Language Models, Graph Query Language, Graph Databases

Contribution Types: NLP engineering experiment

Languages Studied: English, Chinese

Submission Number: 3946