CyVerACT: An Agentic Cypher Translation Workflow over Knowledge Graphs
Year: 2026
Type: article
Abstract: Question Answering (QA) over Knowledge Graphs (KGs) has greatly benefited from the rapid growth of Large Language Models (LLMs), which enable the translation of natural language questions into Cypher queries. Most existing approaches rely on one-shot generation via in-context learning or on fine-tuning LLMs; however, both strategies often struggle to generate accurate or executable queries, particularly when dealing with complex or unfamiliar graph schemas. To address these limitations, in this work we propose CyVerACT, an agentic workflow for Text-to-Cypher generation that empowers LLMs with execution- and schema-aware feedback mechanisms. CyVerACT leverages CyVer, a software tool that evaluates Cypher queries in terms of syntax validity and semantic compliance with respect to a specific KG schema, and detects their points of failure. The system customizes the input graph schema based on the input question and iteratively refines the generated queries taking advantage of the error metadata from CyVer to guide subsequent LLM generations. We evaluated and compared CyVerACT to existing single-shot generation and iterative refinement approaches in two publicly available Text-to-Cypher datasets of 2180 entries across various domains and complexities, using both foundational (e.g., GPT-4o, LLama-3) and fine-tuned state-of-the-art models. Experimental results demonstrate that the proposed workflow significantly improves query correctness and execution success rates, achieving up to 52.7% gain in accuracy in terms of syntax validity and schema access, and 13.5% gain in exact match.
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,670
OpenAlex ID: W7155223385