Text-to-CQL Based on Large Language Model and Graph Pattern Enhancement

With the wide application of knowledge graphs in various scenarios and Neo4j graph databases becoming excellent knowledge graph carriers, Cypher (CQL for short) has become the most popular graph database query language. However, when performing graph database retrieval, the complex pattern and syntax make constructing CQL statements a complicated and time-consuming task. Therefore, similar to Text-to-SQL, it is necessary and urgent to study an effective method for end-to-end transformation of natural languages into CQL. There have been many advances in Text-to-SQL for traditional relational databases, but these methods cannot be well applied to Text-to-CQL, so there is still a lack of effective work on Text-to-CQL for graph databases. In recent years, as large language models have been widely applied to different tasks with good results, and considering that the fundamental difference between CQL and SQL is the representation of graph patterns. We propose in this paper PA-LLM, a Text-to-CQL method that utilizes large language models combined with graph patterns enhancement, which combines large language models and subdivides the graph patterns into three categories according to their respective characteristics. They are simple query patterns, multi-hop query patterns, and function query patterns. For different graph patterns, the method optimizes the process of generating CQL for the model, which can be subdivided into four sub-methods, simple pattern enhancement method, multi-hop pattern enhancement method, function pattern enhancement method, and entity-relationship enhancement method. The experimental results show that the method improves the quality of the CQL statements generated by the model, including the logical accuracy ACCLX and the execution result accuracy ACCEX, and achieves better results on the SpCQL dataset proposed by the National Defense University of Science and Technology (NDUST). It provides an effective solution for the task of converting CQL to natural language end-to-end.