Achieving Precise Text-To-Cypher Via Grounded Knowledge Graph Data Generation

A new synthetic data generation method allows small language models to match much larger proprietary systems on the task of translating natural language into Cypher queries, according to a paper submitted to arXiv on 12 June 2026 [1]. The approach could let organizations deploy accurate, locally run conversational interfaces for property-graph databases without sacrificing data sovereignty or funding expensive manual annotation [1]. Property graphs are increasingly used as database frameworks for heterogeneous data sources, but accessing the information inside them requires conversational interfaces powered by Text-to-Cypher parsers [1]. Large language models can drive these parsers, yet data-sovereignty requirements often force organizations to use smaller, locally deployable models that struggle with zero-shot performance [3]. Fine-tuning those small models demands annotated data, and acquiring it has remained a significant bottleneck [3]. The new method, described in the paper “Achieving Precise Text-To-Cypher Via Grounded Knowledge Graph Data Generation,” automatically generates training data that covers a wider range of Cypher syntax than prior work [3]. Experiments across all major Text-to-Cypher benchmarks show the synthetic data can significantly lift the performance of small LLMs, bringing them close to the accuracy of much larger proprietary models [1][3]. The authors note that because the resulting models are small, the approach can also lead to a lighter environmental footprint [3].

Earlier efforts to tackle the same problem have taken different paths. A 2025 COLING paper introduced SyntheT2C, a methodology that builds synthetic Query-Cypher pairs through LLM-based prompting and template-filling, and demonstrated its effectiveness on two medical knowledge-graph databases [5]. Another pipeline, detailed in a 2024 preprint, generates diverse graph schemas and uses an “LLM-As-Database-Filler” to populate Neo4j instances, retaining only executable queries that return correct results [4]. The new method builds on that lineage by expanding the expressivity of generated queries while ensuring the added complexity does not degrade data quality [3].

The paper appeared on arXiv, the open-access e-print repository that hosts preprints across physics, computer science, and other fields [9]. The abstract page includes links to community-built tools such as the Bibliographic Explorer and Connected Papers, which are part of the arXivLabs framework that lets third-party developers add experimental features to the site [7][8].